diff --git a/bindings/python/src/transcribe_cpp/__init__.py b/bindings/python/src/transcribe_cpp/__init__.py
index b13a92ad..8967a178 100644
--- a/bindings/python/src/transcribe_cpp/__init__.py
+++ b/bindings/python/src/transcribe_cpp/__init__.py
@@ -582,7 +582,7 @@ def _stream_update_from(u) -> StreamUpdate:
 
 
 def _build_run_params(task, language, target_language, timestamps,
-                      keep_special_tags, spec_k_drafts):
+                      keep_special_tags, spec_k_drafts, context=None):
     if not isinstance(spec_k_drafts, int) or spec_k_drafts < -1:
         raise InvalidArgument(
             f"spec_k_drafts must be -1 (family default), 0 (disabled), or a "
@@ -596,6 +596,7 @@ def _build_run_params(task, language, target_language, timestamps,
     params.target_language = target_language.encode("utf-8") if target_language else None
     params.keep_special_tags = keep_special_tags
     params.spec_k_drafts = spec_k_drafts
+    params.context = context.encode("utf-8") if context else None
     return params
 
 
@@ -984,6 +985,7 @@ def run(self, pcm: PCMLike, *, task: Task = "transcribe",
             timestamps: Timestamps = "auto",
             keep_special_tags: bool = False,
             spec_k_drafts: int = -1,
+            context: str | None = None,
             family: FamilyExtension | None = None) -> Result:
         """Transcribe 16 kHz mono float32 PCM and return a materialized Result.
 
@@ -992,6 +994,10 @@ def run(self, pcm: PCMLike, *, task: Task = "transcribe",
         ``spec_k_drafts`` tunes speculative decoding on models whose
         capabilities advertise ``supports_spec_decode`` (-1 = family default,
         0 = disabled, >0 = draft length; silently ignored elsewhere).
+        ``context`` is free text (names, jargon, prior transcript) that
+        biases recognition on models advertising the ``initial_prompt``
+        feature (whisper, qwen3_asr, funasr_nano); the output remains a
+        transcript, and unsupported models warn and ignore it.
 
         On ``Aborted`` (via :meth:`cancel`) and ``OutputTruncated`` the
         partial transcript is preserved and attached to the exception as
@@ -999,7 +1005,7 @@ def run(self, pcm: PCMLike, *, task: Task = "transcribe",
         self._cancel.clear()
         array, n_samples = _pcm_to_carray(pcm)
         params = _build_run_params(task, language, target_language, timestamps,
-                                   keep_special_tags, spec_k_drafts)
+                                   keep_special_tags, spec_k_drafts, context)
         ext = self._resolve_family(family, "run") if family is not None else None
         if ext is not None:
             params.family = ctypes.cast(
@@ -1020,6 +1026,7 @@ def run_batch(self, pcms: Sequence[PCMLike], *, task: Task = "transcribe",
                   timestamps: Timestamps = "auto",
                   keep_special_tags: bool = False,
                   spec_k_drafts: int = -1,
+                  context: str | None = None,
                   family: FamilyExtension | None = None,
                   return_exceptions: bool = False) -> list[Result | TranscribeError]:
         """Transcribe several utterances in one dispatch — one Result each.
@@ -1054,7 +1061,7 @@ def run_batch(self, pcms: Sequence[PCMLike], *, task: Task = "transcribe",
             counts[k] = n
 
         params = _build_run_params(task, language, target_language, timestamps,
-                                   keep_special_tags, spec_k_drafts)
+                                   keep_special_tags, spec_k_drafts, context)
         ext = self._resolve_family(family, "run") if family is not None else None
         if ext is not None:
             params.family = ctypes.cast(
@@ -1107,6 +1114,7 @@ def stream(self, *, task: Task = "transcribe", language: str | None = None,
                target_language: str | None = None, timestamps: Timestamps = "none",
                keep_special_tags: bool = False, commit_policy: CommitPolicy = "auto",
                stable_prefix_agreement_n: int = 0,
+               context: str | None = None,
                family: FamilyExtension | None = None) -> Stream:
         """Begin streaming on this session and return a Stream to feed audio to.
 
@@ -1119,7 +1127,7 @@ def stream(self, *, task: Task = "transcribe", language: str | None = None,
         # spec_k_drafts is an offline-decode knob; streaming always uses the
         # family default (-1).
         run_params = _build_run_params(task, language, target_language, timestamps,
-                                       keep_special_tags, -1)
+                                       keep_special_tags, -1, context)
         sp = _StreamParams()
         _lib.transcribe_stream_params_init(_byref(sp))
         sp.commit_policy = _enum(_COMMIT_POLICIES, commit_policy, "commit_policy")
@@ -1348,6 +1356,7 @@ def transcribe(
     timestamps: Timestamps = "auto",
     keep_special_tags: bool = False,
     spec_k_drafts: int = -1,
+    context: str | None = None,
     family: FamilyExtension | None = None,
 ) -> Result:
     """Transcribe *pcm* in one call and return a materialized Result.
@@ -1362,7 +1371,7 @@ def transcribe(
     session_opts = dict(n_threads=n_threads, kv_type=kv_type, n_ctx=n_ctx)
     run_opts = dict(task=task, language=language, target_language=target_language,
                     timestamps=timestamps, keep_special_tags=keep_special_tags,
-                    spec_k_drafts=spec_k_drafts, family=family)
+                    spec_k_drafts=spec_k_drafts, context=context, family=family)
 
     if isinstance(model, Model):
         with model.session(**session_opts) as session:
diff --git a/bindings/python/src/transcribe_cpp/_generated.py b/bindings/python/src/transcribe_cpp/_generated.py
index 3862bd79..a4b0bb00 100644
--- a/bindings/python/src/transcribe_cpp/_generated.py
+++ b/bindings/python/src/transcribe_cpp/_generated.py
@@ -13,7 +13,7 @@
 # Stable digest of the ABI surface below (structs, enums, macros, layout,
 # prototypes). A native provider package echoes this back so the API
 # package can reject an ABI-mismatched provider before dlopen.
-PUBLIC_HEADER_HASH = "86b16dd97ad1cb58"
+PUBLIC_HEADER_HASH = "3ba8dde028c487b6"
 
 # === enum constants ===
 TRANSCRIBE_OK = 0
@@ -152,7 +152,7 @@ class transcribe_whisper_chunk_trace(_c.Structure):
 transcribe_backend_device._fields_ = [("struct_size", _c.c_uint64), ("name", _c.c_char_p), ("description", _c.c_char_p), ("kind", _c.c_char_p), ("device_id", _c.c_char_p), ("memory_total", _c.c_uint64), ("memory_free", _c.c_uint64), ("device_type", _c.c_int)]
 transcribe_model_load_params._fields_ = [("struct_size", _c.c_uint64), ("backend", _c.c_int), ("gpu_device", _c.c_int)]
 transcribe_session_params._fields_ = [("struct_size", _c.c_uint64), ("n_threads", _c.c_int), ("kv_type", _c.c_int), ("n_ctx", _c.c_int32)]
-transcribe_run_params._fields_ = [("struct_size", _c.c_uint64), ("task", _c.c_int), ("timestamps", _c.c_int), ("pnc", _c.c_int), ("itn", _c.c_int), ("language", _c.c_char_p), ("target_language", _c.c_char_p), ("keep_special_tags", _c.c_bool), ("family", _c.POINTER(transcribe_ext)), ("spec_k_drafts", _c.c_int32)]
+transcribe_run_params._fields_ = [("struct_size", _c.c_uint64), ("task", _c.c_int), ("timestamps", _c.c_int), ("pnc", _c.c_int), ("itn", _c.c_int), ("language", _c.c_char_p), ("target_language", _c.c_char_p), ("keep_special_tags", _c.c_bool), ("family", _c.POINTER(transcribe_ext)), ("spec_k_drafts", _c.c_int32), ("context", _c.c_char_p)]
 transcribe_capabilities._fields_ = [("struct_size", _c.c_uint64), ("native_sample_rate", _c.c_int32), ("n_languages", _c.c_int), ("languages", _c.POINTER(_c.c_char_p)), ("max_timestamp_kind", _c.c_int), ("supports_language_detect", _c.c_bool), ("supports_translate", _c.c_bool), ("supports_streaming", _c.c_bool), ("supports_spec_decode", _c.c_bool), ("max_audio_ms", _c.c_int64), ("n_translate_target_languages", _c.c_int), ("translate_target_languages", _c.POINTER(_c.c_char_p))]
 transcribe_session_limits._fields_ = [("struct_size", _c.c_uint64), ("effective_n_ctx", _c.c_int32), ("effective_max_audio_ms", _c.c_int64), ("max_kv_bytes", _c.c_int64)]
 transcribe_stream_params._fields_ = [("struct_size", _c.c_uint64), ("family", _c.POINTER(transcribe_ext)), ("commit_policy", _c.c_int), ("stable_prefix_agreement_n", _c.c_uint32)]
@@ -194,7 +194,7 @@ class transcribe_whisper_chunk_trace(_c.Structure):
     'transcribe_backend_device': {'size': 64, 'align': 8, 'offsets': {'struct_size': 0, 'name': 8, 'description': 16, 'kind': 24, 'device_id': 32, 'memory_total': 40, 'memory_free': 48, 'device_type': 56}},
     'transcribe_model_load_params': {'size': 16, 'align': 8, 'offsets': {'struct_size': 0, 'backend': 8, 'gpu_device': 12}},
     'transcribe_session_params': {'size': 24, 'align': 8, 'offsets': {'struct_size': 0, 'n_threads': 8, 'kv_type': 12, 'n_ctx': 16}},
-    'transcribe_run_params': {'size': 64, 'align': 8, 'offsets': {'struct_size': 0, 'task': 8, 'timestamps': 12, 'pnc': 16, 'itn': 20, 'language': 24, 'target_language': 32, 'keep_special_tags': 40, 'family': 48, 'spec_k_drafts': 56}},
+    'transcribe_run_params': {'size': 72, 'align': 8, 'offsets': {'struct_size': 0, 'task': 8, 'timestamps': 12, 'pnc': 16, 'itn': 20, 'language': 24, 'target_language': 32, 'keep_special_tags': 40, 'family': 48, 'spec_k_drafts': 56, 'context': 64}},
     'transcribe_capabilities': {'size': 56, 'align': 8, 'offsets': {'struct_size': 0, 'native_sample_rate': 8, 'n_languages': 12, 'languages': 16, 'max_timestamp_kind': 24, 'supports_language_detect': 28, 'supports_translate': 29, 'supports_streaming': 30, 'supports_spec_decode': 31, 'max_audio_ms': 32, 'n_translate_target_languages': 40, 'translate_target_languages': 48}},
     'transcribe_session_limits': {'size': 32, 'align': 8, 'offsets': {'struct_size': 0, 'effective_n_ctx': 8, 'effective_max_audio_ms': 16, 'max_kv_bytes': 24}},
     'transcribe_stream_params': {'size': 24, 'align': 8, 'offsets': {'struct_size': 0, 'family': 8, 'commit_policy': 16, 'stable_prefix_agreement_n': 20}},
diff --git a/bindings/python/uv.lock b/bindings/python/uv.lock
index 5da6ab48..6a4539b0 100644
--- a/bindings/python/uv.lock
+++ b/bindings/python/uv.lock
@@ -386,7 +386,7 @@ wheels = [
 
 [[package]]
 name = "transcribe-cpp"
-version = "0.0.1"
+version = "0.1.1"
 source = { editable = "." }
 
 [package.optional-dependencies]
@@ -402,8 +402,8 @@ test = [
 requires-dist = [
     { name = "numpy", marker = "extra == 'test'" },
     { name = "pytest", marker = "extra == 'test'", specifier = ">=7" },
-    { name = "transcribe-cpp-native", specifier = "==0.0.1.*" },
-    { name = "transcribe-cpp-native-cu12", marker = "extra == 'cu12'", specifier = "==0.0.1.*" },
+    { name = "transcribe-cpp-native", specifier = "==0.1.1.*" },
+    { name = "transcribe-cpp-native-cu12", marker = "extra == 'cu12'", specifier = "==0.1.1.*" },
 ]
 
 [[package]]
diff --git a/bindings/rust/sys/src/transcribe_sys.rs b/bindings/rust/sys/src/transcribe_sys.rs
index d5a353db..8bc68612 100644
--- a/bindings/rust/sys/src/transcribe_sys.rs
+++ b/bindings/rust/sys/src/transcribe_sys.rs
@@ -1,11 +1,11 @@
 // @generated by `cargo xtask bindgen` from include/transcribe/extensions.h
 // DO NOT EDIT BY HAND. Regenerate: `cargo xtask bindgen`.
-// Pinned to include/transcribe.abihash = 86b16dd97ad1cb58
+// Pinned to include/transcribe.abihash = 3ba8dde028c487b6
 
 /// The public-ABI digest these bindings were generated against
 /// (sha256/16 over the normalized FFI surface). The load-time version
 /// gate and the CI drift check both anchor on this value.
-pub const PUBLIC_HEADER_HASH: &str = "86b16dd97ad1cb58";
+pub const PUBLIC_HEADER_HASH: &str = "3ba8dde028c487b6";
 
 /* automatically generated by rust-bindgen 0.72.1 */
 
@@ -329,10 +329,11 @@ pub struct transcribe_run_params {
     pub keep_special_tags: bool,
     pub family: *const transcribe_ext,
     pub spec_k_drafts: i32,
+    pub context: *const ::std::os::raw::c_char,
 }
 #[allow(clippy::unnecessary_operation, clippy::identity_op)]
 const _: () = {
-    ["Size of transcribe_run_params"][::std::mem::size_of::<transcribe_run_params>() - 64usize];
+    ["Size of transcribe_run_params"][::std::mem::size_of::<transcribe_run_params>() - 72usize];
     ["Alignment of transcribe_run_params"]
         [::std::mem::align_of::<transcribe_run_params>() - 8usize];
     ["Offset of field: transcribe_run_params::struct_size"]
@@ -355,6 +356,8 @@ const _: () = {
         [::std::mem::offset_of!(transcribe_run_params, family) - 48usize];
     ["Offset of field: transcribe_run_params::spec_k_drafts"]
         [::std::mem::offset_of!(transcribe_run_params, spec_k_drafts) - 56usize];
+    ["Offset of field: transcribe_run_params::context"]
+        [::std::mem::offset_of!(transcribe_run_params, context) - 64usize];
 };
 unsafe extern "C" {
     pub fn transcribe_run_params_init(params: *mut transcribe_run_params);
diff --git a/bindings/rust/transcribe-cpp/src/session.rs b/bindings/rust/transcribe-cpp/src/session.rs
index e58a9265..f54eb884 100644
--- a/bindings/rust/transcribe-cpp/src/session.rs
+++ b/bindings/rust/transcribe-cpp/src/session.rs
@@ -38,6 +38,11 @@ pub struct RunOptions {
     pub keep_special_tags: bool,
     /// Speculative-decode draft length. `-1` = family default, `0` = disabled.
     pub spec_k_drafts: i32,
+    /// Free-text recognition-biasing context (names, jargon, prior
+    /// transcript). Honored by models advertising the initial-prompt
+    /// feature (whisper, qwen3_asr, funasr_nano); the output remains a
+    /// transcript, and unsupported models warn and ignore it.
+    pub context: Option<String>,
     /// Optional family-specific run extension (e.g. whisper decode knobs).
     pub family: Option<RunExtension>,
 }
@@ -53,6 +58,7 @@ impl Default for RunOptions {
             target_language: None,
             keep_special_tags: false,
             spec_k_drafts: -1,
+            context: None,
             family: None,
         }
     }
@@ -144,7 +150,7 @@ impl Session {
     /// On an aborted or truncated decode the partial transcript is preserved
     /// on the returned [`Error::Aborted`] / [`Error::OutputTruncated`].
     pub fn run(&mut self, pcm: &[f32], options: &RunOptions) -> Result<Transcript> {
-        let (params, _lang, _target, _family) = build_run_params(options)?;
+        let (params, _lang, _target, _context, _family) = build_run_params(options)?;
         let n = clamp_len(pcm.len())?;
 
         // The compute path is serialized per model; hold the lock for the native
@@ -187,7 +193,7 @@ impl Session {
         pcms: &[&[f32]],
         options: &RunOptions,
     ) -> Result<Vec<Result<Transcript>>> {
-        let (params, _lang, _target, _family) = build_run_params(options)?;
+        let (params, _lang, _target, _context, _family) = build_run_params(options)?;
         let ptrs: Vec<*const f32> = pcms.iter().map(|p| p.as_ptr()).collect();
         let lens: Vec<i32> = pcms
             .iter()
@@ -263,7 +269,7 @@ impl Session {
     /// Dropping the returned `Stream` abandons it and returns the session to
     /// idle.
     pub fn stream(&mut self, run: &RunOptions, stream: &StreamOptions) -> Result<Stream<'_>> {
-        let (run_params, _lang, _target, _family) = build_run_params(run)?;
+        let (run_params, _lang, _target, _context, _family) = build_run_params(run)?;
         let (stream_params, _stream_family) = build_stream_params(stream);
         {
             // Claim the model's compute lease for the whole stream lifetime: a
@@ -396,6 +402,7 @@ type RunParamsBundle = (
     sys::transcribe_run_params,
     Option<CString>,
     Option<CString>,
+    Option<CString>,
     Option<RunExtRaw>,
 );
 
@@ -415,8 +422,10 @@ fn build_run_params(o: &RunOptions) -> Result<RunParamsBundle> {
 
     let lang = o.language.as_deref().map(CString::new).transpose()?;
     let target = o.target_language.as_deref().map(CString::new).transpose()?;
+    let context = o.context.as_deref().map(CString::new).transpose()?;
     params.language = lang.as_ref().map_or(std::ptr::null(), |c| c.as_ptr());
     params.target_language = target.as_ref().map_or(std::ptr::null(), |c| c.as_ptr());
+    params.context = context.as_ref().map_or(std::ptr::null(), |c| c.as_ptr());
 
     let family = o
         .family
@@ -425,7 +434,7 @@ fn build_run_params(o: &RunOptions) -> Result<RunParamsBundle> {
         .transpose()?;
     params.family = family.as_ref().map_or(std::ptr::null(), |f| f.ext_ptr());
 
-    Ok((params, lang, target, family))
+    Ok((params, lang, target, context, family))
 }
 
 /// PCM/utterance lengths cross the ABI as `int`; reject anything that overflows.
diff --git a/bindings/swift/Sources/TranscribeCpp/ABIHash.swift b/bindings/swift/Sources/TranscribeCpp/ABIHash.swift
index 0ad79fb9..9792f63d 100644
--- a/bindings/swift/Sources/TranscribeCpp/ABIHash.swift
+++ b/bindings/swift/Sources/TranscribeCpp/ABIHash.swift
@@ -13,7 +13,7 @@ import CTranscribe
 extension Transcribe {
     /// sha256/16 of the normalized public FFI surface, pinned to the value in
     /// include/transcribe.abihash at the time this binding was last reviewed.
-    public static let pinnedHeaderHash = "86b16dd97ad1cb58"
+    public static let pinnedHeaderHash = "3ba8dde028c487b6"
 
     /// The public-ABI digest this binding was reviewed against (16 hex chars).
     public static func headerHash() -> String { pinnedHeaderHash }
diff --git a/bindings/swift/Sources/TranscribeCpp/Options.swift b/bindings/swift/Sources/TranscribeCpp/Options.swift
index 4401fda3..ee064005 100644
--- a/bindings/swift/Sources/TranscribeCpp/Options.swift
+++ b/bindings/swift/Sources/TranscribeCpp/Options.swift
@@ -121,6 +121,11 @@ public struct RunOptions: Sendable {
     public var keepSpecialTags: Bool
     /// Speculative-decode draft length: -1 = family default, 0 = disabled.
     public var specKDrafts: Int32
+    /// Free-text recognition-biasing context (names, jargon, prior
+    /// transcript). Honored by models advertising the initial-prompt
+    /// feature (whisper, qwen3_asr, funasr_nano); the output remains a
+    /// transcript, and unsupported models warn and ignore it.
+    public var context: String?
     /// Family-specific run extension (whisper run options); M3.
     public var family: RunExtension?
 
@@ -136,6 +141,7 @@ public struct RunOptions: Sendable {
         targetLanguage: String? = nil,
         keepSpecialTags: Bool = false,
         specKDrafts: Int32 = -1,
+        context: String? = nil,
         family: RunExtension? = nil
     ) {
         self.task = task
@@ -146,6 +152,7 @@ public struct RunOptions: Sendable {
         self.targetLanguage = targetLanguage
         self.keepSpecialTags = keepSpecialTags
         self.specKDrafts = specKDrafts
+        self.context = context
         self.family = family
     }
 
@@ -165,9 +172,12 @@ public struct RunOptions: Sendable {
             params.language = lang
             return try withOptionalCString(targetLanguage) { tgt in
                 params.target_language = tgt
-                return try withRunExtension(family) { ext in
-                    params.family = ext
-                    return try withUnsafePointer(to: &params) { try body($0) }
+                return try withOptionalCString(context) { ctx in
+                    params.context = ctx
+                    return try withRunExtension(family) { ext in
+                        params.family = ext
+                        return try withUnsafePointer(to: &params) { try body($0) }
+                    }
                 }
             }
         }
diff --git a/bindings/typescript/src/_generated.ts b/bindings/typescript/src/_generated.ts
index bf1da0d9..e2800941 100644
--- a/bindings/typescript/src/_generated.ts
+++ b/bindings/typescript/src/_generated.ts
@@ -11,7 +11,7 @@
 // Stable digest of the ABI surface (structs, enums, macros, layout,
 // prototypes), computed by the Python oracle and pinned here so a header
 // ABI change turns this binding's drift check red for conscious review.
-export const PUBLIC_HEADER_HASH = "86b16dd97ad1cb58";
+export const PUBLIC_HEADER_HASH = "3ba8dde028c487b6";
 
 // === enum constants ===
 export const TRANSCRIBE_OK = 0;
@@ -110,7 +110,7 @@ export const STRUCT_LAYOUT: Record<string, StructLayout> = {
   'transcribe_backend_device': { size: 64, align: 8, offsets: {'struct_size': 0, 'name': 8, 'description': 16, 'kind': 24, 'device_id': 32, 'memory_total': 40, 'memory_free': 48, 'device_type': 56} },
   'transcribe_model_load_params': { size: 16, align: 8, offsets: {'struct_size': 0, 'backend': 8, 'gpu_device': 12} },
   'transcribe_session_params': { size: 24, align: 8, offsets: {'struct_size': 0, 'n_threads': 8, 'kv_type': 12, 'n_ctx': 16} },
-  'transcribe_run_params': { size: 64, align: 8, offsets: {'struct_size': 0, 'task': 8, 'timestamps': 12, 'pnc': 16, 'itn': 20, 'language': 24, 'target_language': 32, 'keep_special_tags': 40, 'family': 48, 'spec_k_drafts': 56} },
+  'transcribe_run_params': { size: 72, align: 8, offsets: {'struct_size': 0, 'task': 8, 'timestamps': 12, 'pnc': 16, 'itn': 20, 'language': 24, 'target_language': 32, 'keep_special_tags': 40, 'family': 48, 'spec_k_drafts': 56, 'context': 64} },
   'transcribe_capabilities': { size: 56, align: 8, offsets: {'struct_size': 0, 'native_sample_rate': 8, 'n_languages': 12, 'languages': 16, 'max_timestamp_kind': 24, 'supports_language_detect': 28, 'supports_translate': 29, 'supports_streaming': 30, 'supports_spec_decode': 31, 'max_audio_ms': 32, 'n_translate_target_languages': 40, 'translate_target_languages': 48} },
   'transcribe_session_limits': { size: 32, align: 8, offsets: {'struct_size': 0, 'effective_n_ctx': 8, 'effective_max_audio_ms': 16, 'max_kv_bytes': 24} },
   'transcribe_stream_params': { size: 24, align: 8, offsets: {'struct_size': 0, 'family': 8, 'commit_policy': 16, 'stable_prefix_agreement_n': 20} },
@@ -152,7 +152,7 @@ export function defineTypes(koffi: any): Record<string, any> {
   T['transcribe_backend_device'] = koffi.struct({ struct_size: 'uint64_t', name: 'char *', description: 'char *', kind: 'char *', device_id: 'char *', memory_total: 'uint64_t', memory_free: 'uint64_t', device_type: 'int' });
   T['transcribe_model_load_params'] = koffi.struct({ struct_size: 'uint64_t', backend: 'int', gpu_device: 'int' });
   T['transcribe_session_params'] = koffi.struct({ struct_size: 'uint64_t', n_threads: 'int', kv_type: 'int', n_ctx: 'int32_t' });
-  T['transcribe_run_params'] = koffi.struct({ struct_size: 'uint64_t', task: 'int', timestamps: 'int', pnc: 'int', itn: 'int', language: 'char *', target_language: 'char *', keep_special_tags: 'bool', family: 'void *', spec_k_drafts: 'int32_t' });
+  T['transcribe_run_params'] = koffi.struct({ struct_size: 'uint64_t', task: 'int', timestamps: 'int', pnc: 'int', itn: 'int', language: 'char *', target_language: 'char *', keep_special_tags: 'bool', family: 'void *', spec_k_drafts: 'int32_t', context: 'char *' });
   T['transcribe_capabilities'] = koffi.struct({ struct_size: 'uint64_t', native_sample_rate: 'int32_t', n_languages: 'int', languages: 'void *', max_timestamp_kind: 'int', supports_language_detect: 'bool', supports_translate: 'bool', supports_streaming: 'bool', supports_spec_decode: 'bool', max_audio_ms: 'int64_t', n_translate_target_languages: 'int', translate_target_languages: 'void *' });
   T['transcribe_session_limits'] = koffi.struct({ struct_size: 'uint64_t', effective_n_ctx: 'int32_t', effective_max_audio_ms: 'int64_t', max_kv_bytes: 'int64_t' });
   T['transcribe_stream_params'] = koffi.struct({ struct_size: 'uint64_t', family: 'void *', commit_policy: 'int', stable_prefix_agreement_n: 'uint32_t' });
diff --git a/bindings/typescript/src/index.ts b/bindings/typescript/src/index.ts
index d87440f2..48582b6c 100644
--- a/bindings/typescript/src/index.ts
+++ b/bindings/typescript/src/index.ts
@@ -776,6 +776,7 @@ export class Session {
     if (opts.keepSpecialTags !== undefined)
       p.keep_special_tags = opts.keepSpecialTags;
     if (opts.specKDrafts !== undefined) p.spec_k_drafts = opts.specKDrafts;
+    if (opts.context !== undefined) p.context = opts.context;
     if (opts.family)
       p.family = buildFamily(n, this.#model.handle, opts.family, "run");
     return p;
@@ -872,6 +873,7 @@ export class Session {
       timestamps: opts.timestamps,
       keepSpecialTags: opts.keepSpecialTags,
       specKDrafts: -1,
+      context: opts.context,
     });
     const sp: any = {};
     F.streamParamsInit(sp);
diff --git a/bindings/typescript/src/types.ts b/bindings/typescript/src/types.ts
index 12d65131..5ee7d03f 100644
--- a/bindings/typescript/src/types.ts
+++ b/bindings/typescript/src/types.ts
@@ -136,6 +136,11 @@ export interface TranscribeOptions {
   keepSpecialTags?: boolean;
   /** Speculative-decode draft count; -1 = family default. */
   specKDrafts?: number;
+  /** Free-text recognition-biasing context (names, jargon, prior
+   *  transcript). Honored by models advertising the initial-prompt
+   *  feature (whisper, qwen3_asr, funasr_nano); the output remains a
+   *  transcript, and unsupported models warn and ignore it. */
+  context?: string;
   /** Cancel the run cooperatively. */
   signal?: AbortSignal;
   /** A run-slot family extension (e.g. whisper). */
@@ -181,6 +186,8 @@ export interface StreamOptions {
   targetLanguage?: string;
   timestamps?: TimestampKind;
   keepSpecialTags?: boolean;
+  /** Free-text recognition-biasing context; see TranscribeOptions.context. */
+  context?: string;
   commitPolicy?: CommitPolicy;
   stablePrefixAgreementN?: number;
   /** A stream-slot family extension (moonshine, parakeet, voxtral). */
diff --git a/docs/models/fun-asr-nano.md b/docs/models/fun-asr-nano.md
index 07beb665..eeffc114 100644
--- a/docs/models/fun-asr-nano.md
+++ b/docs/models/fun-asr-nano.md
@@ -84,6 +84,12 @@ All Fun-ASR-Nano variants support:
   punctuation) via `--itn` on the CLI, or
   `transcribe_funasr_nano_params { use_itn = true }` via the library
   API.
+- **Hotword / context biasing** via `--context` on the CLI or
+  `transcribe_run_params::context` in the API. The text fills the
+  hotword slot of the trained prompt template (upstream
+  `FunASRNano.get_prompt` hotwords path), so comma-separated
+  keyword lists match the trained shape best — e.g.
+  `--context "pull buoy, catch-up freestyle"`.
 
 What's not supported (consistent across the family): translation,
 real-time streaming, long-form chunking, timestamps, VAD, speaker
diff --git a/docs/models/qwen3-asr.md b/docs/models/qwen3-asr.md
index a5553c73..54adcc08 100644
--- a/docs/models/qwen3-asr.md
+++ b/docs/models/qwen3-asr.md
@@ -72,6 +72,12 @@ All Qwen3-ASR variants support:
 
 - **Transcription** of 16 kHz mono WAV input.
 - **Auto language detection** across 30 languages.
+- **Context biasing** — Qwen3-ASR's trained "customized recognition"
+  channel. Pass names/jargon/prior context via `--context` (or
+  `transcribe_run_params::context`); the text rides the chat template's
+  system slot, exactly where the upstream reference puts its `context`
+  string. Keyword lists and free paragraphs both work; the output
+  remains a transcript of the audio.
 
 What's not supported (consistent across the family): translation,
 real-time streaming, VAD, speaker diarization, timestamps. See the
diff --git a/docs/models/voxtral.md b/docs/models/voxtral.md
index eb8e4435..38c4fd92 100644
--- a/docs/models/voxtral.md
+++ b/docs/models/voxtral.md
@@ -42,5 +42,10 @@ reports the exact per-session value. See the
 - The 24B is the larger sibling — same architecture, scaled decoder. It is a
   GPU-class model (BF16/F16 need ~50 GB) and should be run at **batch size
   ≤ 8**; see its [card](voxtral-small-24b-2507.md) for details.
+- **No context biasing**: Voxtral's transcription template has no slot for
+  `transcribe_run_params::context` (a supplied context warns and is
+  ignored). The model's free-text instruct mode exists upstream but changes
+  the task rather than biasing recognition, so it is not mapped onto the
+  core context field; if exposed later it would be a voxtral run extension.
 - Upstream: [`mistralai/Voxtral-Mini-3B-2507`](https://huggingface.co/mistralai/Voxtral-Mini-3B-2507),
   [`mistralai/Voxtral-Small-24B-2507`](https://huggingface.co/mistralai/Voxtral-Small-24B-2507).
diff --git a/docs/models/whisper.md b/docs/models/whisper.md
index cb940ded..52e02470 100644
--- a/docs/models/whisper.md
+++ b/docs/models/whisper.md
@@ -91,6 +91,11 @@ section.
 All Whisper variants support:
 
 - **Transcription** of 16 kHz mono WAV input.
+- **Context biasing** — pass names/jargon/prior-transcript text via
+  `--context` (or `transcribe_run_params::context`); it is injected as
+  Whisper previous-context after `<|startofprev|>`. The whisper run
+  extension's `initial_prompt` / `prompt_tokens` are the power-user
+  form of the same mechanism and win when both are set.
 - **Long-form audio** via 30-second chunked decoding with the
   prev-context window assembly described in the family doc.
 - **Segment timestamps** — the finest granularity the library emits for
diff --git a/examples/cli/main.cpp b/examples/cli/main.cpp
index c407f98b..8db3cd1f 100644
--- a/examples/cli/main.cpp
+++ b/examples/cli/main.cpp
@@ -131,6 +131,10 @@ struct cli_args {
     int                        gpu_device   = 0;  // --device N: 0 = auto, >0 = registry index
     transcribe_timestamp_kind  timestamps   = TRANSCRIBE_TIMESTAMPS_AUTO;
 
+    // Generic recognition-biasing context (transcribe_run_params::context).
+    // Families without the INITIAL_PROMPT feature warn and ignore it.
+    std::string context;  // --context TEXT
+
     // Whisper-family knobs. Ignored for non-Whisper models.
     std::string                              initial_prompt;                    // --initial-prompt TEXT
     bool                                     whisper_set              = false;
@@ -211,7 +215,11 @@ void print_usage(const char * argv0) {
                  "  --batch-jsonl         output one JSON line per file (for batch)\n"
                  "  --batch-size N        group N utterances into one transcribe_run_batch\n"
                  "                        call (offline only; 0/1 = per-file serial loop)\n"
-                 "  --initial-prompt TEXT (whisper) initial prompt text for context biasing\n"
+                 "  --context TEXT        recognition-biasing context (names, jargon,\n"
+                 "                        prior transcript); models without the feature\n"
+                 "                        warn and ignore it\n"
+                 "  --initial-prompt TEXT (whisper) initial prompt text for context biasing;\n"
+                 "                        wins over --context on whisper\n"
                  "  --temperature F       (whisper) tier-0 sampling temperature (default 0 = greedy)\n"
                  "  --condition-on-prev-tokens (whisper) carry prev-chunk tokens across chunks\n"
                  "  --prompt-condition T  (whisper) prompt placement: first|all (default: first)\n"
@@ -439,6 +447,12 @@ bool parse_args(int argc, char ** argv, cli_args & out) {
                 std::fprintf(stderr, "error: --batch-size must be >= 0\n");
                 return false;
             }
+        } else if (a == "--context") {
+            const char * v = take_value(a.c_str());
+            if (!v) {
+                return false;
+            }
+            out.context = v;
         } else if (a == "--initial-prompt") {
             const char * v = take_value(a.c_str());
             if (!v) {
@@ -698,6 +712,9 @@ int main(int argc, char ** argv) {
         if (args.canary_pnc_set) {
             rp.pnc = args.canary_pnc ? TRANSCRIBE_PNC_MODE_ON : TRANSCRIBE_PNC_MODE_OFF;
         }
+        if (!args.context.empty()) {
+            rp.context = args.context.c_str();
+        }
 
         // Whisper run extension. Allocated outside rp's scope so its
         // bytes outlive the per-file loop below; the library copies
@@ -1075,6 +1092,9 @@ int main(int argc, char ** argv) {
         if (args.canary_pnc_set) {
             rp.pnc = args.canary_pnc ? TRANSCRIBE_PNC_MODE_ON : TRANSCRIBE_PNC_MODE_OFF;
         }
+        if (!args.context.empty()) {
+            rp.context = args.context.c_str();
+        }
 
         struct transcribe_whisper_run_ext wx;
         transcribe_whisper_run_ext_init(&wx);
diff --git a/include/transcribe.abihash b/include/transcribe.abihash
index dc4df617..bbb0b5aa 100644
--- a/include/transcribe.abihash
+++ b/include/transcribe.abihash
@@ -1 +1 @@
-86b16dd97ad1cb58
+3ba8dde028c487b6
diff --git a/include/transcribe.h b/include/transcribe.h
index 81b2d37b..ab39dbab 100644
--- a/include/transcribe.h
+++ b/include/transcribe.h
@@ -984,14 +984,14 @@ TRANSCRIBE_API void transcribe_session_params_init(struct transcribe_session_par
  *
  * target_language: target language for translation tasks, or NULL.
  *
- * String-pointer lifetime (language / target_language): caller-owned, and
- * the library copies what it needs before the API call returns. This holds
- * for transcribe_run / transcribe_run_batch (synchronous) AND for
- * transcribe_stream_begin: the dispatcher copies these strings into
- * session-owned storage at begin, so the caller may free its params —
- * including the strings — the moment begin returns, even though the stream
- * keeps running. Bindings rely on this: ctypes/FFI-managed string buffers
- * die when the begin wrapper returns.
+ * String-pointer lifetime (language / target_language / context):
+ * caller-owned, and the library copies what it needs before the API call
+ * returns. This holds for transcribe_run / transcribe_run_batch
+ * (synchronous) AND for transcribe_stream_begin: the dispatcher copies
+ * these strings into session-owned storage at begin, so the caller may
+ * free its params — including the strings — the moment begin returns,
+ * even though the stream keeps running. Bindings rely on this:
+ * ctypes/FFI-managed string buffers die when the begin wrapper returns.
  *
  * keep_special_tags: keep special vocabulary tags (e.g. <|...|>) in the
  *                     returned text fields. Default (false) strips them
@@ -1044,6 +1044,43 @@ struct transcribe_run_params {
      *   to know whether the field will take effect.
      */
     int32_t spec_k_drafts;
+
+    /*
+     * context: free-text that conditions recognition — names, jargon,
+     *   hotwords, prior-transcript or domain context. NULL (default) or
+     *   empty means no context.
+     *
+     *   Contract: the output remains a faithful transcript of the audio.
+     *   context biases WHAT the model hears (spelling of names, choice
+     *   between homophones), never what it does — it is not an
+     *   instruction channel, and families never route it anywhere that
+     *   would change the task. Each family maps the string to its native
+     *   mechanism:
+     *
+     *     whisper      previous-context prefix after <|startofprev|>
+     *                  (same path as transcribe_whisper_run_ext's
+     *                  initial_prompt; the ext, when present, wins — see
+     *                  include/transcribe/whisper.h for precedence).
+     *     qwen3_asr    system-message context slot (the model's
+     *                  trained biasing channel; accepts keyword lists
+     *                  or free paragraphs).
+     *     funasr_nano  hotword/context slot of the trained prompt
+     *                  template. Keyword-list-style text ("Foo, Bar,
+     *                  Baz") matches the trained shape best.
+     *
+     *   A non-empty context against a model for which
+     *   transcribe_model_supports(model, TRANSCRIBE_FEATURE_INITIAL_PROMPT)
+     *   is false emits a WARN and proceeds without it (same advisory
+     *   semantics as pnc/itn).
+     *
+     *   Field-promotion rule (why this is core and e.g. voxtral's
+     *   instruct mode is not): a knob enters transcribe_run_params when
+     *   two or more families can honor the same caller intent WITHOUT
+     *   changing the output contract. Anything that changes what the
+     *   model does (free-form task instructions, chat) stays on the
+     *   owning family's extension header.
+     */
+    const char * context;
 };
 
 TRANSCRIBE_API void transcribe_run_params_init(struct transcribe_run_params * params);
@@ -1231,9 +1268,13 @@ TRANSCRIBE_API transcribe_status transcribe_model_get_capabilities(const struct
  *
  * Feature meanings:
  *
- *   INITIAL_PROMPT       The model accepts a free-text or token
- *                        prompt to bias decoding. Today: whisper
- *                        only; reached via transcribe_whisper_run_ext.
+ *   INITIAL_PROMPT       The model honors transcribe_run_params::context
+ *                        — free text that biases recognition while the
+ *                        output remains a transcript. Today: whisper,
+ *                        qwen3_asr, funasr_nano. Whisper additionally
+ *                        accepts token-level prompts and decode knobs
+ *                        via transcribe_whisper_run_ext, which takes
+ *                        precedence over the core field.
  *
  *   TEMPERATURE_FALLBACK The model runs a multi-tier temperature loop
  *                        with metric-driven fallback. Today: whisper.
diff --git a/include/transcribe/whisper.h b/include/transcribe/whisper.h
index f40d0d66..c5884c58 100644
--- a/include/transcribe/whisper.h
+++ b/include/transcribe/whisper.h
@@ -102,6 +102,11 @@ struct transcribe_whisper_run_ext {
      *       rejected with TRANSCRIBE_ERR_INVALID_ARG, mirroring HF's
      *       own check.
      *
+     *   Else: the core transcribe_run_params::context string, when set,
+     *       feeds this same path (identical tokenization and special-token
+     *       rejection). The ext fields, being the more specific surface,
+     *       always win over the core field.
+     *
      *   Else: no initial prompt.
      */
     const char *    initial_prompt;
diff --git a/src/arch/funasr_nano/capabilities.cpp b/src/arch/funasr_nano/capabilities.cpp
index b3c163c7..f52ac995 100644
--- a/src/arch/funasr_nano/capabilities.cpp
+++ b/src/arch/funasr_nano/capabilities.cpp
@@ -24,6 +24,11 @@ void apply_family_invariants(transcribe_model & model) {
     // runtime toggle.
     transcribe::set_feature(&model, TRANSCRIBE_FEATURE_CANCELLATION, true);
     transcribe::set_feature(&model, TRANSCRIBE_FEATURE_ITN, true);
+
+    // transcribe_run_params::context is honored: the string fills the
+    // hotword/context slot of the trained prompt template (the reference's
+    // FunASRNano.get_prompt hotwords path — 热词列表：[...]).
+    transcribe::set_feature(&model, TRANSCRIBE_FEATURE_INITIAL_PROMPT, true);
 }
 
 }  // namespace transcribe::funasr_nano
diff --git a/src/arch/funasr_nano/model.cpp b/src/arch/funasr_nano/model.cpp
index c1aa74c1..0e4460cb 100644
--- a/src/arch/funasr_nano/model.cpp
+++ b/src/arch/funasr_nano/model.cpp
@@ -147,15 +147,49 @@ transcribe_status resolve_chat_tokens(const transcribe::Tokenizer & tok, ChatTok
 }
 
 // Build the language/itn prompt text that the reference's
-// FunASRNano.get_prompt produces. hotwords-empty path only.
-std::string build_funasr_prompt_text(const char * lang, bool use_itn) {
+// FunASRNano.get_prompt produces. A non-empty `context` renders the
+// reference's hotwords path — the context/hotword preamble with the
+// caller's text in the 热词列表 ("hotword list") slot, exactly where
+// get_prompt puts its ", ".join(hotwords):
+//
+//   请结合上下文信息，更加准确地完成语音转写任务。
+//   如果没有相关信息，我们会留空。\n\n\n**上下文信息：**\n\n\n
+//   热词列表：[{context}]\n
+std::string build_funasr_prompt_text(const char * context, const char * lang, bool use_itn) {
     std::string out;
+    if (context != nullptr && context[0] != '\0') {
+        // 请结合上下文信息，更加准确地完成语音转写任务。
+        // 如果没有相关信息，我们会留空。
+        out =
+            "\xE8\xAF\xB7\xE7\xBB\x93\xE5\x90\x88\xE4\xB8\x8A"
+            "\xE4\xB8\x8B\xE6\x96\x87\xE4\xBF\xA1\xE6\x81\xAF"
+            "\xEF\xBC\x8C\xE6\x9B\xB4\xE5\x8A\xA0\xE5\x87\x86"
+            "\xE7\xA1\xAE\xE5\x9C\xB0\xE5\xAE\x8C\xE6\x88\x90"
+            "\xE8\xAF\xAD\xE9\x9F\xB3\xE8\xBD\xAC\xE5\x86\x99"
+            "\xE4\xBB\xBB\xE5\x8A\xA1\xE3\x80\x82\xE5\xA6\x82"
+            "\xE6\x9E\x9C\xE6\xB2\xA1\xE6\x9C\x89\xE7\x9B\xB8"
+            "\xE5\x85\xB3\xE4\xBF\xA1\xE6\x81\xAF\xEF\xBC\x8C"
+            "\xE6\x88\x91\xE4\xBB\xAC\xE4\xBC\x9A\xE7\x95\x99"
+            "\xE7\xA9\xBA\xE3\x80\x82";
+        out += "\n\n\n";
+        // **上下文信息：**
+        out +=
+            "\x2A\x2A\xE4\xB8\x8A\xE4\xB8\x8B\xE6\x96\x87\xE4\xBF\xA1"
+            "\xE6\x81\xAF\xEF\xBC\x9A\x2A\x2A";
+        out += "\n\n\n";
+        // 热词列表：[
+        out +=
+            "\xE7\x83\xAD\xE8\xAF\x8D\xE5\x88\x97\xE8\xA1\xA8"
+            "\xEF\xBC\x9A\x5B";
+        out += context;
+        out += "]\n";
+    }
     if (lang != nullptr && lang[0] != '\0') {
         // 语音转写成 = "transcribe to" / "transcribe into"
-        out = "\xE8\xAF\xAD\xE9\x9F\xB3\xE8\xBD\xAC\xE5\x86\x99\xE6\x88\x90";
+        out += "\xE8\xAF\xAD\xE9\x9F\xB3\xE8\xBD\xAC\xE5\x86\x99\xE6\x88\x90";
         out += lang;
     } else {
-        out = "\xE8\xAF\xAD\xE9\x9F\xB3\xE8\xBD\xAC\xE5\x86\x99";
+        out += "\xE8\xAF\xAD\xE9\x9F\xB3\xE8\xBD\xAC\xE5\x86\x99";
     }
     if (!use_itn) {
         // ，不进行文本规整 = "; do not apply text normalization"
@@ -227,6 +261,7 @@ transcribe_status encode_with_chat_specials(const transcribe::Tokenizer & tok,
 // <|im_start|>/<|im_end|> within each segment.
 transcribe_status build_funasr_nano_prompt(const transcribe::Tokenizer & tok,
                                            const ChatTokens &            ct,
+                                           const char *                  context,
                                            const char *                  language,
                                            bool                          use_itn,
                                            int                           fake_token_len,
@@ -235,7 +270,19 @@ transcribe_status build_funasr_nano_prompt(const transcribe::Tokenizer & tok,
     out_ids.clear();
     out_fbank_beg = 0;
 
-    const std::string prompt_text = build_funasr_prompt_text(language, use_itn);
+    // Caller context rides inside the user turn that
+    // encode_with_chat_specials scans for <|im_start|>/<|im_end|> markers;
+    // reject those literals in context text so caller data cannot inject
+    // control tokens (mirrors whisper's special-token rejection).
+    if (context != nullptr &&
+        (std::strstr(context, "<|im_start|>") != nullptr || std::strstr(context, "<|im_end|>") != nullptr)) {
+        log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
+                "funasr_nano: context text must not contain chat control "
+                "markers (<|im_start|>/<|im_end|>)");
+        return TRANSCRIBE_ERR_INVALID_ARG;
+    }
+
+    const std::string prompt_text = build_funasr_prompt_text(context, language, use_itn);
 
     std::string seg_a =
         "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
@@ -658,7 +705,8 @@ transcribe_status run(transcribe_session *          session,
     std::vector<int32_t> prompt_ids;
     int                  fbank_beg = 0;
     if (const transcribe_status st =
-            build_funasr_nano_prompt(cm->tok, cm->chat_tokens, lang, use_itn, fake_token_len, prompt_ids, fbank_beg);
+            build_funasr_nano_prompt(cm->tok, cm->chat_tokens, transcribe_run_params_context(params), lang, use_itn,
+                                     fake_token_len, prompt_ids, fbank_beg);
         st != TRANSCRIBE_OK) {
         return st;
     }
@@ -1139,8 +1187,8 @@ transcribe_status run_batch(transcribe_session *          session,
             continue;
         }
         int fbank_beg = 0;
-        if (build_funasr_nano_prompt(cm->tok, cm->chat_tokens, lang, use_itn, T_audio[b], prompt_ids[b], fbank_beg) !=
-            TRANSCRIBE_OK) {
+        if (build_funasr_nano_prompt(cm->tok, cm->chat_tokens, transcribe_run_params_context(params), lang, use_itn,
+                                     T_audio[b], prompt_ids[b], fbank_beg) != TRANSCRIBE_OK) {
             continue;
         }
         T_prompt[b] = static_cast<int>(prompt_ids[b].size());
diff --git a/src/arch/qwen3_asr/capabilities.cpp b/src/arch/qwen3_asr/capabilities.cpp
index 4737e588..14dc7e58 100644
--- a/src/arch/qwen3_asr/capabilities.cpp
+++ b/src/arch/qwen3_asr/capabilities.cpp
@@ -20,6 +20,11 @@ void apply_family_invariants(transcribe_model & model) {
     // Cancellation is wired at the per-run level. No PNC/ITN toggle; the
     // Whisper-specific features do not apply here.
     transcribe::set_feature(&model, TRANSCRIBE_FEATURE_CANCELLATION, true);
+
+    // transcribe_run_params::context is honored: the string rides the chat
+    // template's system slot, Qwen3-ASR's trained biasing channel (the
+    // reference's TranscribeArgs context / _build_messages system message).
+    transcribe::set_feature(&model, TRANSCRIBE_FEATURE_INITIAL_PROMPT, true);
 }
 
 }  // namespace transcribe::qwen3_asr
diff --git a/src/arch/qwen3_asr/model.cpp b/src/arch/qwen3_asr/model.cpp
index 6e382fbe..06bd3304 100644
--- a/src/arch/qwen3_asr/model.cpp
+++ b/src/arch/qwen3_asr/model.cpp
@@ -364,16 +364,21 @@ transcribe_status resolve_chat_tokens(const transcribe::Tokenizer & tok, ChatTok
 // Build the prompt token sequence + audio-position list, mirroring the
 // Qwen3-ASR chat template at the token level:
 //
-//   <|im_start|>system\n<|im_end|>\n
+//   <|im_start|>system\n{context}<|im_end|>\n
 //   <|im_start|>user\n<|audio_start|><|audio_pad|>*T_enc<|audio_end|><|im_end|>\n
 //   <|im_start|>assistant\n[language {Name}<asr_text>]?
 //
-// System prompt is empty. A non-null `lang_prefix_ids` (resolved via
+// The system slot carries the caller's recognition-biasing context verbatim
+// (the reference builds {"role": "system", "content": context or ""} —
+// qwen_asr/inference/qwen3_asr.py:_build_messages); a null/empty
+// `context_ids` leaves it empty. A non-null `lang_prefix_ids` (resolved via
 // encode_language_prefix) is appended after the trailing newline to force an
-// output language; kept out of here so this stays a pure token-id assembler.
+// output language; both are kept out of here so this stays a pure token-id
+// assembler.
 void build_prompt_tokens(const QwenAsrHParams &       hp,
                          const ChatTokens &           ct,
                          int                          T_enc,
+                         const std::vector<int32_t> * context_ids,
                          const std::vector<int32_t> * lang_prefix_ids,
                          std::vector<int32_t> &       out_ids,
                          std::vector<int64_t> &       out_audio_positions) {
@@ -383,6 +388,9 @@ void build_prompt_tokens(const QwenAsrHParams &       hp,
     out_ids.push_back(ct.im_start);
     out_ids.push_back(ct.role_system);
     out_ids.push_back(ct.newline);
+    if (context_ids != nullptr && !context_ids->empty()) {
+        out_ids.insert(out_ids.end(), context_ids->begin(), context_ids->end());
+    }
     out_ids.push_back(ct.im_end);
     out_ids.push_back(ct.newline);
 
@@ -564,6 +572,19 @@ transcribe_status run(transcribe_session *          session,
         lang_prefix_ptr = &lang_prefix_ids;
     }
 
+    // Recognition-biasing context -> system-message tokens (the reference's
+    // TranscribeArgs context / _build_messages system slot). Plain BPE; the
+    // template's control tokens are pushed by build_prompt_tokens itself.
+    std::vector<int32_t>         context_ids;
+    const std::vector<int32_t> * context_ptr = nullptr;
+    if (const char * ctx_text = transcribe_run_params_context(params); ctx_text != nullptr) {
+        if (const transcribe_status st = cm->tok.encode(ctx_text, context_ids); st != TRANSCRIBE_OK) {
+            transcribe::log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "qwen3_asr run: tokenizer.encode failed on context");
+            return st;
+        }
+        context_ptr = &context_ids;
+    }
+
     transcribe::debug::init();
 
     // Mel front-end.
@@ -725,7 +746,7 @@ transcribe_status run(transcribe_session *          session,
     // Prompt construction.
     std::vector<int32_t> prompt_ids;
     std::vector<int64_t> audio_positions;
-    build_prompt_tokens(cm->hparams, cm->chat_tokens, T_enc, lang_prefix_ptr, prompt_ids, audio_positions);
+    build_prompt_tokens(cm->hparams, cm->chat_tokens, T_enc, context_ptr, lang_prefix_ptr, prompt_ids, audio_positions);
     const int T_prompt   = static_cast<int>(prompt_ids.size());
     const int prefix_len = audio_positions.empty() ? 0 : static_cast<int>(audio_positions.front());
     const int suffix_len = T_prompt - prefix_len - T_enc;
@@ -1528,6 +1549,18 @@ transcribe_status run_batch(transcribe_session *          session,
         lang_prefix_ptr = &lang_prefix_ids;
     }
 
+    // Shared recognition-biasing context (same one-run_params contract);
+    // encoded once, injected into every utterance's system slot.
+    std::vector<int32_t>         context_ids;
+    const std::vector<int32_t> * context_ptr = nullptr;
+    if (const char * ctx_text = transcribe_run_params_context(params); ctx_text != nullptr) {
+        if (cm->tok.encode(ctx_text, context_ids) != TRANSCRIBE_OK) {
+            transcribe::log_msg(TRANSCRIBE_LOG_LEVEL_ERROR, "qwen3_asr run_batch: tokenizer.encode failed on context");
+            return TRANSCRIBE_ERR_GGUF;
+        }
+        context_ptr = &context_ids;
+    }
+
     // Pass 1: per-utterance encoder + prefill into KV slabs.
     std::vector<std::vector<int32_t>> generated(n);
     std::vector<int>                  T_prompt(n, 0);
@@ -1563,7 +1596,7 @@ transcribe_status run_batch(transcribe_session *          session,
             continue;
         }
         std::vector<int64_t> ap;
-        build_prompt_tokens(cm->hparams, cm->chat_tokens, T_enc[b], lang_prefix_ptr, prompt_ids[b], ap);
+        build_prompt_tokens(cm->hparams, cm->chat_tokens, T_enc[b], context_ptr, lang_prefix_ptr, prompt_ids[b], ap);
         T_prompt[b] = static_cast<int>(prompt_ids[b].size());
         prefix_len  = ap.empty() ? 0 : static_cast<int>(ap.front());
         // Same gate as single-shot run(); the rest of the batch still runs.
diff --git a/src/arch/voxtral/capabilities.cpp b/src/arch/voxtral/capabilities.cpp
index 58e69108..e5eded12 100644
--- a/src/arch/voxtral/capabilities.cpp
+++ b/src/arch/voxtral/capabilities.cpp
@@ -22,10 +22,13 @@ void apply_family_invariants(transcribe_model & model) {
     // by read_capability_kv at load; default it true here as a fallback.
     caps.supports_translate = true;
 
-    // Per-run cancellation; the free-text prompt is wired via the
-    // INITIAL_PROMPT feature on the Voxtral run extension.
+    // Per-run cancellation only. INITIAL_PROMPT stays false: Voxtral's
+    // transcription template has no recognition-biasing slot for
+    // transcribe_run_params::context — the only place free text can go is
+    // instruct mode, which replaces the [TRANSCRIBE] task (a different
+    // output contract). If free-text instruct is ever exposed, it belongs
+    // on a voxtral run extension, not the core context field.
     transcribe::set_feature(&model, TRANSCRIBE_FEATURE_CANCELLATION, true);
-    transcribe::set_feature(&model, TRANSCRIBE_FEATURE_INITIAL_PROMPT, true);
 }
 
 }  // namespace transcribe::voxtral
diff --git a/src/arch/whisper/model.cpp b/src/arch/whisper/model.cpp
index e8c6d711..79cb76e0 100644
--- a/src/arch/whisper/model.cpp
+++ b/src/arch/whisper/model.cpp
@@ -1476,8 +1476,16 @@ transcribe_status whisper_run(transcribe_session *          session,
     if (prev_sot_id < 0) {
         prev_sot_id = cm->tok.find("<|startofprev|>");
     }
-    const bool prompt_requested = (wp->prompt_tokens != nullptr && wp->n_prompt_tokens > 0) ||
-                                  (wp->initial_prompt != nullptr && wp->initial_prompt[0] != '\0');
+    // Effective initial-prompt string. The whisper run ext wins when it
+    // carries one; otherwise the core run_params::context maps onto the same
+    // previous-context mechanism (transcribe_run_params_context is size-gated
+    // and returns nullptr for absent/empty, so this is non-null iff some
+    // caller supplied non-empty prompt text).
+    const char * initial_prompt_text = (wp->initial_prompt != nullptr && wp->initial_prompt[0] != '\0') ?
+                                           wp->initial_prompt :
+                                           transcribe_run_params_context(params);
+    const bool   prompt_requested =
+        (wp->prompt_tokens != nullptr && wp->n_prompt_tokens > 0) || initial_prompt_text != nullptr;
     if (prev_sot_id < 0 && (prompt_requested || wp->condition_on_prev_tokens)) {
         log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                 "whisper run: model has no <|startofprev|> token; "
@@ -1487,9 +1495,10 @@ transcribe_status whisper_run(transcribe_session *          session,
 
     // Resolve initial prompt -> text-only token ids (library prepends
     // <|startofprev|>). Two paths: prompt_tokens (caller-owned, verbatim,
-    // text-side only); or initial_prompt string, tokenized as HF's
-    // get_prompt_ids form ("<|startofprev|> " + strip) with any special token
-    // (id >= eos_id) in the text rejected (tokenization_whisper.py).
+    // text-side only); or the effective prompt string (ext initial_prompt or
+    // core context), tokenized as HF's get_prompt_ids form
+    // ("<|startofprev|> " + strip) with any special token (id >= eos_id) in
+    // the text rejected (tokenization_whisper.py).
     const int max_prev_cap =
         wp->max_prev_context_tokens > 0 ? wp->max_prev_context_tokens : (cm->hparams.dec_max_target_positions / 2 - 1);
     std::vector<int32_t> prompt_text_ids;
@@ -1505,8 +1514,8 @@ transcribe_status whisper_run(transcribe_session *          session,
             return TRANSCRIBE_ERR_INVALID_ARG;
         }
         prompt_text_ids.assign(wp->prompt_tokens, wp->prompt_tokens + wp->n_prompt_tokens);
-    } else if (wp->initial_prompt != nullptr && wp->initial_prompt[0] != '\0') {
-        std::string s(wp->initial_prompt);
+    } else if (initial_prompt_text != nullptr) {
+        std::string s(initial_prompt_text);
         auto        issp = [](char c) {
             return c == ' ' || c == '\t' || c == '\n' || c == '\r';
         };
@@ -1540,7 +1549,7 @@ transcribe_status whisper_run(transcribe_session *          session,
                         const int    id    = cm->tok.find(piece);
                         if (id >= eos_id) {
                             log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
-                                    "whisper run: initial_prompt contains "
+                                    "whisper run: initial_prompt/context contains "
                                     "disallowed special token \"%s\" (id %d)",
                                     piece.c_str(), id);
                             return TRANSCRIBE_ERR_INVALID_ARG;
@@ -1555,13 +1564,13 @@ transcribe_status whisper_run(transcribe_session *          session,
             if (cm->tok.encode(text, prompt_text_ids) != TRANSCRIBE_OK) {
                 log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
                         "whisper run: tokenizer.encode failed on "
-                        "initial_prompt");
+                        "initial_prompt/context");
                 return TRANSCRIBE_ERR_GGUF;
             }
             for (int32_t id : prompt_text_ids) {
                 if (id >= eos_id) {
                     log_msg(TRANSCRIBE_LOG_LEVEL_ERROR,
-                            "whisper run: initial_prompt contains "
+                            "whisper run: initial_prompt/context contains "
                             "disallowed special token id %d",
                             id);
                     return TRANSCRIBE_ERR_INVALID_ARG;
@@ -2576,8 +2585,13 @@ transcribe_status whisper_run_batch(transcribe_session *          session,
     if (prev_sot_id < 0) {
         prev_sot_id = cm->tok.find("<|startofprev|>");
     }
-    const bool           prompt_requested = (wp->prompt_tokens != nullptr && wp->n_prompt_tokens > 0) ||
-                                            (wp->initial_prompt != nullptr && wp->initial_prompt[0] != '\0');
+    // Effective prompt string mirrors the serial path: ext initial_prompt
+    // wins, core run_params::context is the fallback.
+    const char * initial_prompt_text = (wp->initial_prompt != nullptr && wp->initial_prompt[0] != '\0') ?
+                                           wp->initial_prompt :
+                                           transcribe_run_params_context(params);
+    const bool   prompt_requested =
+        (wp->prompt_tokens != nullptr && wp->n_prompt_tokens > 0) || initial_prompt_text != nullptr;
     std::vector<int32_t> prev_tokens;
     if (prompt_requested) {
         if (prev_sot_id < 0) {
@@ -2592,7 +2606,7 @@ transcribe_status whisper_run_batch(transcribe_session *          session,
             }
             ptext.assign(wp->prompt_tokens, wp->prompt_tokens + wp->n_prompt_tokens);
         } else {
-            std::string s(wp->initial_prompt);
+            std::string s(initial_prompt_text);
             auto        issp = [](char c) {
                 return c == ' ' || c == '\t' || c == '\n' || c == '\r';
             };
diff --git a/src/transcribe-session.h b/src/transcribe-session.h
index 7d769eaf..c730562b 100644
--- a/src/transcribe-session.h
+++ b/src/transcribe-session.h
@@ -38,6 +38,25 @@ inline int32_t transcribe_session_params_n_ctx(const struct transcribe_session_p
     return params->n_ctx;
 }
 
+// Read transcribe_run_params::context with the same struct_size guard:
+// context is a trailing field (appended after spec_k_drafts), so an older
+// caller's smaller struct may not include it. Returns nullptr for a NULL
+// params, a pre-context struct, or an empty string, so families need only
+// one non-null check to mean "caller supplied context text".
+inline const char * transcribe_run_params_context(const struct transcribe_run_params * params) {
+    if (params == nullptr) {
+        return nullptr;
+    }
+    const size_t field_end = offsetof(struct transcribe_run_params, context) + sizeof(params->context);
+    if (params->struct_size < static_cast<uint64_t>(field_end)) {
+        return nullptr;
+    }
+    if (params->context == nullptr || params->context[0] == '\0') {
+        return nullptr;
+    }
+    return params->context;
+}
+
 struct transcribe_session {
     // The model this session was constructed from. Borrowed pointer:
     // the caller is required (per the public threading contract) to keep
@@ -241,6 +260,7 @@ struct transcribe_session {
     // next begin mutates them.
     std::string stream_language_owned;
     std::string stream_target_language_owned;
+    std::string stream_context_owned;
 
     // UI-facing streaming text state. `full_text` above remains the raw
     // model hypothesis. `stream_committed_text` is the append-only public
diff --git a/src/transcribe.cpp b/src/transcribe.cpp
index 6e254815..ecd132aa 100644
--- a/src/transcribe.cpp
+++ b/src/transcribe.cpp
@@ -1351,6 +1351,18 @@ void warn_unsupported_advisory(const struct transcribe_model * model, const stru
                       req, arch_name, variant);
         transcribe_log_emit_or_stderr(TRANSCRIBE_LOG_LEVEL_WARN, buf);
     }
+    // transcribe_run_params_context is size-gated, so a pre-context caller
+    // struct reads as "no context" here rather than as garbage bytes.
+    if (transcribe_run_params_context(rp) != nullptr &&
+        !transcribe::has_feature(model, TRANSCRIBE_FEATURE_INITIAL_PROMPT)) {
+        std::snprintf(buf, sizeof(buf),
+                      "transcribe_run: caller supplied context text but model '%s' "
+                      "(variant '%s') has no recognition-biasing mechanism; the "
+                      "context is ignored. Use transcribe_model_supports(model, "
+                      "TRANSCRIBE_FEATURE_INITIAL_PROMPT) to pre-check.",
+                      arch_name, variant);
+        transcribe_log_emit_or_stderr(TRANSCRIBE_LOG_LEVEL_WARN, buf);
+    }
 }
 
 }  // namespace
@@ -1771,6 +1783,12 @@ static transcribe_status transcribe_stream_begin_impl(struct transcribe_session
     // wants a run-slot ext at stream begin must plumb it deliberately.
     session->stream_language_owned        = run_params->language != nullptr ? run_params->language : "";
     session->stream_target_language_owned = run_params->target_language != nullptr ? run_params->target_language : "";
+    // context is size-gated (trailing field): the accessor reads it only
+    // when the caller's struct is new enough to carry it.
+    {
+        const char * ctx_in           = transcribe_run_params_context(run_params);
+        session->stream_context_owned = ctx_in != nullptr ? ctx_in : "";
+    }
     // PREFIX copy, not struct assignment: the size gate above admits any
     // struct_size >= k_min_run_params_size, so a conforming caller's
     // allocation may be SHORTER than sizeof (fields past `family`, e.g.
@@ -1785,7 +1803,8 @@ static transcribe_status transcribe_stream_begin_impl(struct transcribe_session
     run_params_owned.language = run_params->language != nullptr ? session->stream_language_owned.c_str() : nullptr;
     run_params_owned.target_language =
         run_params->target_language != nullptr ? session->stream_target_language_owned.c_str() : nullptr;
-    run_params_owned.family = nullptr;
+    run_params_owned.context = !session->stream_context_owned.empty() ? session->stream_context_owned.c_str() : nullptr;
+    run_params_owned.family  = nullptr;
 
     const transcribe_status st = session->model->arch->stream_begin(session, &run_params_owned, stream_params);
     if (st != TRANSCRIBE_OK) {
diff --git a/tests/api_smoke.c b/tests/api_smoke.c
index 04ae026b..5abf103c 100644
--- a/tests/api_smoke.c
+++ b/tests/api_smoke.c
@@ -234,6 +234,8 @@ static void test_init_macros(void) {
     CHECK(rp_macro.target_language == NULL);
     CHECK(rp_macro.keep_special_tags == false);
     CHECK(rp_macro.family == NULL);
+    CHECK(rp_macro.spec_k_drafts == -1);
+    CHECK(rp_macro.context == NULL);
 
     struct transcribe_stream_params sp_macro;
     transcribe_stream_params_init(&sp_macro);
diff --git a/tests/qwen3_asr_smoke.cpp b/tests/qwen3_asr_smoke.cpp
index 98ac5ba2..4a61c3b1 100644
--- a/tests/qwen3_asr_smoke.cpp
+++ b/tests/qwen3_asr_smoke.cpp
@@ -169,6 +169,9 @@ int main() {
         CHECK_EQ_INT(caps->n_languages, 4);
         CHECK(caps->languages != nullptr);
         CHECK_EQ_INT(caps->max_timestamp_kind, TRANSCRIBE_TIMESTAMPS_NONE);
+        // run_params::context is honored (system-slot biasing) — the
+        // family invariant advertises it via the feature probe.
+        CHECK(transcribe_model_supports(model, TRANSCRIBE_FEATURE_INITIAL_PROMPT) == true);
     }
 
     // Internal-view assertions.