mlx-bun 0.4.0 → 0.5.0

package/src/vision/qwen3vl-prompt.ts

@@ -18,10 +18,12 @@ import type { ChatMessage, ChatTemplate, ToolDefinition } from "../chat-template
 import type { LoadedTokenizer } from "../tokenizer";
 import type { Qwen35Model } from "../model/qwen3_5";
 import { qwenRopeIndex, type MropeRequestState } from "../model/qwen3-mrope";
-import { preprocessQwen3VLImage, preprocessQwen3VLVideoFrames, pyFixed1 } from "./qwen3vl-preprocess";
+import { preprocessQwen3VLImage, preprocessQwen3VLVideoFrames, pyFixed1, type Qwen3VLPreprocessed } from "./qwen3vl-preprocess";
 import { extractVideoFrames } from "./video-frames";
 import { QWEN3VL_MERGE_SIZE } from "./qwen3vl-preprocess";
 import type { Qwen3VLVisionTower } from "./qwen3vl-tower";
+import { createHash } from "node:crypto";
+import type { EncoderCache } from "../backends/mlx/encoder-cache";
 
 export interface Qwen3VLTokenIds {
   imageTokenId: number; // <|image_pad|>
@@ -34,6 +36,7 @@ export interface Qwen3VLTokenIds {
 
 export interface Qwen3VLVisionPrompt {
   ids: number[];
+  prefixIdentity?: string;
   /** [1, L, H] bf16 spliced input embeddings (caller owns → generate()). */
   embeddings: MlxArray;
   /** Request mRoPE state for Qwen35Model.mrope. */
@@ -50,6 +53,8 @@ export async function buildQwen3VLVisionPrompt(
   tokenIds: Qwen3VLTokenIds,
   renderOptions: Parameters<ChatTemplate["render"]>[1] = {},
   videos: Uint8Array[] = [],
+  nativeWork: <T>(work: () => Promise<T>) => Promise<T> = work => work(),
+  encoderCache?: EncoderCache,
 ): Promise<Qwen3VLVisionPrompt> {
   // The request's full template options (tools, enableThinking,
   // reasoningEffort, preserveThinking) flow through — a media prompt must
@@ -71,9 +76,9 @@ export async function buildQwen3VLVisionPrompt(
 
   // Preprocess each medium in its own appearance order; videos decode to
   // sampled frames through the AVFoundation sidecar first.
-  const imagePps = [];
+  const imagePps: Qwen3VLPreprocessed[] = [];
   for (const img of images) imagePps.push(await preprocessQwen3VLImage(img));
-  const videoPps = [];
+  const videoPps: Qwen3VLPreprocessed[] = [];
   for (const vid of videos)
     videoPps.push(preprocessQwen3VLVideoFrames(await extractVideoFrames(vid)));
 
@@ -93,11 +98,13 @@ export async function buildQwen3VLVisionPrompt(
   const grids: [number, number, number][] = [];
   let imgIdx = 0;
   let vidIdx = 0;
+  let lastMediaEnd = 0;
   for (const t of rawIds) {
     if (t === tokenIds.imageTokenId) {
       const pp = imagePps[imgIdx++]!;
       grids.push(pp.gridThw);
       for (let i = 0; i < pp.imageTokens; i++) ids.push(tokenIds.imageTokenId);
+      lastMediaEnd = ids.length;
     } else if (t === tokenIds.videoTokenId) {
       const pp = videoPps[vidIdx++]!;
       const [gridT, gridH, gridW] = pp.gridThw;
@@ -111,6 +118,7 @@ export async function buildQwen3VLVisionPrompt(
         for (let i = 0; i < frameTokens; i++) ids.push(tokenIds.videoTokenId);
         ids.push(visionEndId);
       }
+      lastMediaEnd = ids.length;
     } else ids.push(t);
   }
 
@@ -126,71 +134,100 @@ export async function buildQwen3VLVisionPrompt(
   // path can't reach any of it (2026-08-18 review: repeated failing media
   // requests would otherwise strand un-admitted GPU memory on the GC
   // backstop).
-  const idsArr = ops.fromInt32(ids, [1, ids.length]);
-  const textEmb = model.embed.encode(idsArr); // [1, L, H]
-  idsArr.dispose();
-  const owned: MlxArray[] = [];
+  // Look up exact preprocessed content before entering the execution queue.
+  // URLs are not identities: their contents can change. Position/timestamp
+  // assembly above remains request-owned even when encoder features are reused.
+  const inputs = [...imagePps, ...videoPps];
+  const keys = encoderCache ? inputs.map(pp => createHash("sha256")
+    .update(`qwen3vl-v1:${JSON.stringify(pp.gridThw)}:${pp.rows}:${pp.cols}:`)
+    .update(new Uint8Array(pp.pixelValues.buffer, pp.pixelValues.byteOffset, pp.pixelValues.byteLength))
+    .digest("hex")) : [];
+  const prefixIdentity = encoderCache && lastMediaEnd > 0
+    ? createHash("sha256").update("qwen-prepared-prefix-v1").update(encoderCache.identity)
+      .update(JSON.stringify({ keys, tokens: ids.slice(0, lastMediaEnd), delta: mrope.delta,
+        positions: mrope.positions.map(axis => Array.from(axis.slice(0, lastMediaEnd))),
+      })).digest("hex") : undefined;
+  const reused: Array<MlxArray | null> = [];
   try {
-    const H = textEmb.shape[2]!;
-    const segments: MlxArray[] = [];
-    let cursor = 0;
-    imgIdx = 0;
-    vidIdx = 0;
-    // A video's tower features cover ALL its frame groups; consecutive pad
-    // RUNS consume consecutive feature-row windows.
-    let vidFeat: MlxArray | null = null;
-    let vidRow = 0;
-    for (let i = 0; i < ids.length; ) {
-      const isImage = ids[i] === tokenIds.imageTokenId;
-      const isVideo = ids[i] === tokenIds.videoTokenId;
-      if (isImage || isVideo) {
-        // Measure this contiguous pad run.
-        let runEnd = i;
-        while (runEnd < ids.length && ids[runEnd] === ids[i]) runEnd++;
-        const runLen = runEnd - i;
-        if (cursor < i) {
-          const seg = textEmb.slice([0, cursor, 0], [1, i, H]);
+    for (const key of keys) reused.push(await encoderCache!.take(key));
+    return await nativeWork(async () => {
+      const encode = (index: number): MlxArray => {
+        const cached = reused[index];
+        if (cached) return cached.slice([0, 0], [...cached.shape]);
+        const features = tower.encode(inputs[index]!);
+        try { if (encoderCache) encoderCache.put(keys[index]!, features); }
+        catch (error) { features.dispose(); throw error; }
+        return features;
+      };
+      const idsArr = ops.fromInt32(ids, [1, ids.length]);
+      const textEmb = model.embed.encode(idsArr); // [1, L, H]
+      idsArr.dispose();
+      const owned: MlxArray[] = [];
+      try {
+        const H = textEmb.shape[2]!;
+        const segments: MlxArray[] = [];
+        let cursor = 0;
+        imgIdx = 0;
+        vidIdx = 0;
+        // A video's tower features cover ALL its frame groups; consecutive pad
+        // RUNS consume consecutive feature-row windows.
+        let vidFeat: MlxArray | null = null;
+        let vidRow = 0;
+        for (let i = 0; i < ids.length; ) {
+          const isImage = ids[i] === tokenIds.imageTokenId;
+          const isVideo = ids[i] === tokenIds.videoTokenId;
+          if (isImage || isVideo) {
+            // Measure this contiguous pad run.
+            let runEnd = i;
+            while (runEnd < ids.length && ids[runEnd] === ids[i]) runEnd++;
+            const runLen = runEnd - i;
+            if (cursor < i) {
+              const seg = textEmb.slice([0, cursor, 0], [1, i, H]);
+              segments.push(seg);
+              owned.push(seg);
+            }
+            let rows: MlxArray;
+            if (isImage) {
+              const feat = encode(imgIdx++); // [runLen, H]
+              owned.push(feat);
+              rows = feat;
+            } else {
+              if (!vidFeat) {
+                vidFeat = encode(imagePps.length + vidIdx); // [imageTokens, H]
+                owned.push(vidFeat);
+                vidRow = 0;
+              }
+              const sl = vidFeat.slice([vidRow, 0], [vidRow + runLen, H]);
+              owned.push(sl);
+              rows = sl;
+              vidRow += runLen;
+              if (vidRow >= videoPps[vidIdx]!.imageTokens) {
+                vidFeat = null;
+                vidIdx++;
+              }
+            }
+            const feat3 = ops.reshape(rows, [1, runLen, H]);
+            segments.push(feat3);
+            owned.push(feat3);
+            cursor = runEnd;
+            i = cursor;
+          } else i++;
+        }
+        if (cursor < ids.length) {
+          const seg = textEmb.slice([0, cursor, 0], [1, ids.length, H]);
           segments.push(seg);
           owned.push(seg);
         }
-        let rows: MlxArray;
-        if (isImage) {
-          const feat = tower.encode(imagePps[imgIdx++]!); // [runLen, H]
-          owned.push(feat);
-          rows = feat;
-        } else {
-          if (!vidFeat) {
-            vidFeat = tower.encode(videoPps[vidIdx]!); // [imageTokens, H]
-            owned.push(vidFeat);
-            vidRow = 0;
-          }
-          const sl = vidFeat.slice([vidRow, 0], [vidRow + runLen, H]);
-          owned.push(sl);
-          rows = sl;
-          vidRow += runLen;
-          if (vidRow >= videoPps[vidIdx]!.imageTokens) {
-            vidFeat = null;
-            vidIdx++;
-          }
-        }
-        const feat3 = ops.reshape(rows, [1, runLen, H]);
-        segments.push(feat3);
-        owned.push(feat3);
-        cursor = runEnd;
-        i = cursor;
-      } else i++;
-    }
-    if (cursor < ids.length) {
-      const seg = textEmb.slice([0, cursor, 0], [1, ids.length, H]);
-      segments.push(seg);
-      owned.push(seg);
-    }
-    const embeddings = segments.length === 1
-      ? ops.contiguous(segments[0]!)
-      : ops.concatAxis(segments, 1);
-    return { ids, embeddings, mrope };
+        const embeddings = segments.length === 1
+          ? ops.contiguous(segments[0]!)
+          : ops.concatAxis(segments, 1);
+        return { ids, embeddings, mrope, prefixIdentity };
+      } finally {
+        for (const s of owned) s.dispose();
+        textEmb.dispose();
+      }
+    });
   } finally {
-    for (const s of owned) s.dispose();
-    textEmb.dispose();
+    for (const features of reused) features?.dispose();
   }
 }

package/src/vision/qwen3vl-tower.ts

@@ -18,6 +18,7 @@
 
 import { ptr, read } from "bun:ffi";
 import { existsSync, readFileSync } from "node:fs";
+import { tensorFingerprint } from "../model/fingerprint";
 import { MlxArray, cpuStream } from "../mlx/array";
 import { C, Dtype } from "../mlx/ffi";
 import * as ops from "../mlx/ops";
@@ -55,6 +56,15 @@ interface Block {
 }
 
 export class Qwen3VLVisionTower {
+  #identity: string | undefined;
+  readonly #identityWeights: Array<[string, MlxArray]>;
+
+  /** Computed once, under native preparation. Exact weight content prevents
+   * cross-artifact SSD reuse even when model aliases and shapes are identical. */
+  get cacheIdentity(): string {
+    if (this.#identity) return this.#identity;
+    return this.#identity = tensorFingerprint(this.#identityWeights);
+  }
   #patchW: MlxArray; // [1536, 1152] — flattened conv, already transposed
   #patchB: MlxArray;
   #posEmbed: MlxArray; // [2304, 1152]
@@ -64,6 +74,7 @@ export class Qwen3VLVisionTower {
   #mergerFc2W: MlxArray; #mergerFc2B: MlxArray;
 
   private constructor(w: Map<string, MlxArray>) {
+    this.#identityWeights = [...w];
     const t = (n: string): MlxArray => {
       const a = w.get(`vision_tower.${n}`);
       if (!a) throw new Error(`vision sidecar missing tensor vision_tower.${n}`);

package/src/vision/siglip.ts

@@ -54,6 +54,7 @@
 // composition order. Left at tier-a for now (good enough — grounded, exact ids,
 // greedy prefix); revisit to drive the encoder to 0% by aligning the op order.
 
+import { tensorFingerprint } from "../model/fingerprint";
 import { ptr, read } from "bun:ffi";
 import { MlxArray, cpuStream } from "../mlx/array";
 import { C, Dtype } from "../mlx/ffi";
@@ -189,6 +190,13 @@ export async function preprocessSiglip(
 
 export class SiglipVisionTower {
   #weights = new Map<string, MlxArray>();
+  #cacheIdentity?: string;
+  /** Immutable encoder weights and computation settings, hashed once on use. */
+  get cacheIdentity(): string {
+    return this.#cacheIdentity ??= new Bun.CryptoHasher("sha256")
+      .update("gemma-siglip-v1").update(JSON.stringify([this.cfg, this.embedScale]))
+      .update(tensorFingerprint(this.#weights)).digest("hex");
+  }
   readonly cfg: SiglipVisionConfig;
   readonly embedScale: number;
   #wdtype: Dtype = Dtype.bfloat16;

package/src/web/app.html

@@ -1118,6 +1118,13 @@
   #chat-attach-btn{flex:0 0 auto;width:50px;height:50px;border-radius:14px;background:var(--card);
     border:1px solid var(--hairline);color:var(--ink);font-size:23px;line-height:1;cursor:pointer;transition:background .15s,border-color .15s}
   #chat-attach-btn:hover{background:var(--card-hover);border-color:rgba(10,132,255,.5)}
+  /* hold-to-talk mic (voice.ts) */
+  #chat-mic-btn{flex:0 0 auto;width:50px;height:50px;border-radius:14px;background:var(--card);
+    border:1px solid var(--hairline);color:var(--ink);font-size:20px;line-height:1;cursor:pointer;user-select:none;-webkit-user-select:none;transition:background .15s,border-color .15s}
+  #chat-mic-btn:hover{background:var(--card-hover);border-color:rgba(10,132,255,.5)}
+  #chat-mic-btn.recording{background:rgba(255,69,58,.18);border-color:rgba(255,69,58,.7);animation:micpulse 1s ease-in-out infinite}
+  #chat-mic-btn.busy{opacity:.6;cursor:progress}
+  @keyframes micpulse{0%,100%{box-shadow:0 0 0 0 rgba(255,69,58,.35)}50%{box-shadow:0 0 0 8px rgba(255,69,58,0)}}
   .attach-row{display:none;flex-wrap:wrap;gap:8px;max-width:840px;width:100%;margin:0 auto 8px}
   .attach-chip{display:flex;align-items:center;gap:7px;background:var(--card);border:1px solid var(--hairline);
     border-radius:10px;padding:5px 9px;font-size:12px;max-width:240px}
@@ -1496,6 +1503,11 @@
         <div class="inputrow">
           <button id="chat-attach-btn" class="iconbtn" title="Attach files or images" aria-label="Attach files">＋</button>
           <input type="file" id="chat-file-input" multiple style="display:none">
+          <!-- Hold-to-talk dictation (voice.ts): shown only when the `ready`
+               frame carries transcription=true (a Whisper checkpoint serves
+               /v1/audio/sessions). Press and hold (mouse/touch, or Space/Enter
+               while focused); release to transcribe into the box. -->
+          <button id="chat-mic-btn" class="iconbtn" title="Hold to talk (Whisper)" aria-label="Hold to talk" style="display:none">🎙</button>
           <div id="chat-box-wrap">
             <!-- Keep the placeholder SHORT: the Enter/Shift+Enter AND "#"
                  hints live in the chat-hint status line below (and the Cmd+/

package/src/web/app.js

@@ -1209,6 +1209,7 @@ function readText(file) {
 class ComposerState {
   attachments = [];
   visionCapable = false;
+  transcriptionCapable = false;
   thinkingCapable = false;
   thinkingOn = true;
   attachSeq = 0;
@@ -1321,6 +1322,9 @@ function updateAttachHint(state) {
   const btn = $("chat-attach-btn");
   if (btn)
     btn.title = state.visionCapable ? "Attach files or images" : "Attach files (this model can't see images)";
+  const mic = $("chat-mic-btn");
+  if (mic)
+    mic.style.display = state.transcriptionCapable ? "" : "none";
 }
 function escHtml2(s) {
   return String(s ?? "").replace(/[&<>"]/g, (c) => ({ "&": "&amp;", "<": "&lt;", ">": "&gt;", '"': "&quot;" })[c]);
@@ -1983,6 +1987,160 @@ function renderQueue(q) {
   bar.style.display = parts.length ? "flex" : "none";
 }
 
+// src/web/src/voice.ts
+var SAMPLE_RATE = 16000;
+var CHUNK_SECONDS = 0.25;
+var take = null;
+var starting = false;
+async function postChunk(id, pcm) {
+  const r = await fetch(`/v1/audio/sessions/${id}/audio`, {
+    method: "POST",
+    headers: { "content-type": "audio/pcm;rate=16000" },
+    body: new Uint8Array(pcm.buffer.slice(pcm.byteOffset, pcm.byteOffset + pcm.byteLength))
+  });
+  if (!r.ok)
+    throw new Error(`audio chunk rejected (${r.status})`);
+}
+function flush(t) {
+  if (t.pendingLen === 0)
+    return;
+  const buf = new Float32Array(t.pendingLen);
+  let o = 0;
+  for (const p of t.pending) {
+    buf.set(p, o);
+    o += p.length;
+  }
+  t.pending = [];
+  t.pendingLen = 0;
+  t.sending = t.sending.then(() => postChunk(t.id, buf)).catch((e) => {
+    t.failed = e.message;
+  });
+}
+async function begin(btn) {
+  if (take || starting)
+    return;
+  starting = true;
+  btn.classList.add("busy");
+  try {
+    const r = await fetch("/v1/audio/sessions", {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({ vad: true, vad_min_speech_ms: 120, temperature: 0 })
+    });
+    if (!r.ok)
+      throw new Error(r.status === 503 ? "no Whisper model on this server" : `session failed (${r.status})`);
+    const { id } = await r.json();
+    const stream = await navigator.mediaDevices.getUserMedia({ audio: { channelCount: 1, echoCancellation: true, noiseSuppression: true } });
+    const ctx = new AudioContext({ sampleRate: SAMPLE_RATE });
+    const src = ctx.createMediaStreamSource(stream);
+    const node = ctx.createScriptProcessor(4096, 1, 1);
+    const t = { id, ctx, stream, node, pending: [], pendingLen: 0, sending: Promise.resolve(), failed: null };
+    node.onaudioprocess = (e) => {
+      if (take !== t)
+        return;
+      const data = e.inputBuffer.getChannelData(0);
+      t.pending.push(new Float32Array(data));
+      t.pendingLen += data.length;
+      if (t.pendingLen >= SAMPLE_RATE * CHUNK_SECONDS)
+        flush(t);
+    };
+    src.connect(node);
+    node.connect(ctx.destination);
+    take = t;
+    btn.classList.add("recording");
+  } catch (e) {
+    toast(`Voice input failed: ${e.message}`, "err");
+  } finally {
+    starting = false;
+    btn.classList.remove("busy");
+  }
+}
+async function end(btn, box) {
+  const t = take;
+  if (!t)
+    return;
+  take = null;
+  btn.classList.remove("recording");
+  btn.classList.add("busy");
+  try {
+    t.node.disconnect();
+    for (const track of t.stream.getTracks())
+      track.stop();
+    t.ctx.close();
+    flush(t);
+    await t.sending;
+    if (t.failed)
+      throw new Error(t.failed);
+    const r = await fetch(`/v1/audio/sessions/${t.id}/finish`, { method: "POST" });
+    if (!r.ok)
+      throw new Error(`transcription failed (${r.status})`);
+    const body = await r.json();
+    const text = (body.text || "").trim();
+    if (!text) {
+      toast(body.mlx_bun?.vad?.speech === false ? "No speech detected" : "Nothing transcribed", "");
+      return;
+    }
+    insertAtCaret(box, text);
+  } catch (e) {
+    toast(`Voice input failed: ${e.message}`, "err");
+  } finally {
+    btn.classList.remove("busy");
+  }
+}
+function insertAtCaret(box, text) {
+  const start = box.selectionStart ?? box.value.length;
+  const endPos = box.selectionEnd ?? start;
+  const before = box.value.slice(0, start);
+  const after = box.value.slice(endPos);
+  const lead = before.length && !/\s$/.test(before) ? " " : "";
+  const trail = after.length && !/^\s/.test(after) ? " " : "";
+  box.value = before + lead + text + trail + after;
+  const caret = (before + lead + text + trail).length;
+  box.setSelectionRange(caret, caret);
+  box.dispatchEvent(new Event("input", { bubbles: true }));
+  box.focus();
+}
+function initVoiceInput(btn, box) {
+  if (!btn)
+    return;
+  if (!navigator.mediaDevices?.getUserMedia) {
+    btn.title = "Microphone capture needs a secure context (https or localhost)";
+  }
+  btn.addEventListener("pointerdown", (e) => {
+    e.preventDefault();
+    btn.setPointerCapture(e.pointerId);
+    begin(btn);
+  });
+  const release = () => {
+    end(btn, box);
+  };
+  btn.addEventListener("pointerup", release);
+  btn.addEventListener("pointercancel", release);
+  btn.addEventListener("lostpointercapture", release);
+  let keyHeld = false;
+  btn.addEventListener("keydown", (e) => {
+    if ((e.key === " " || e.key === "Enter") && !keyHeld) {
+      e.preventDefault();
+      keyHeld = true;
+      begin(btn);
+    }
+  });
+  btn.addEventListener("keyup", (e) => {
+    if ((e.key === " " || e.key === "Enter") && keyHeld) {
+      e.preventDefault();
+      keyHeld = false;
+      release();
+    }
+  });
+  btn.addEventListener("blur", () => {
+    if (keyHeld) {
+      keyHeld = false;
+      release();
+    }
+  });
+  $.voice = { begin: () => begin(btn), end: () => end(btn, box) };
+}
+
 // src/web/src/sessions.ts
 var SESSIONS_STYLE_ID = "mlxbun-sessions-style";
 function injectSessionsStyles() {
@@ -3809,6 +3967,7 @@ function createChatController() {
       case "ready":
         $("nav-model").textContent = m.model || $("nav-model").textContent || "no model";
         composer.visionCapable = !!m.vision;
+        composer.transcriptionCapable = !!m.transcription;
         updateAttachHint(composer);
         composer.thinkingCapable = !!m.thinking;
         updateThinkingToggle(composer);
@@ -4026,6 +4185,7 @@ function createChatController() {
       });
       initMemoryPanel();
       $("chat-attach-btn").onclick = () => $("chat-file-input").click();
+      initVoiceInput($("chat-mic-btn"), box);
       $("chat-file-input").addEventListener("change", (e) => {
         addFiles(composer, [...e.target.files || []]);
         e.target.value = "";

package/src/web/src/chat.ts

@@ -30,6 +30,7 @@ import {
   renderAttachments, renderContext, renderLane, renderQueue, updateAttachHint,
   updateThinkingToggle, type ContextFrame,
 } from "./composer";
+import { initVoiceInput } from "./voice";
 import {
   addMsgActions, newSiblingInfo, newSidebarState, renderSessions, renderSiblingToggle,
   switchSiblingTarget, applySessSearch, type SidebarState, type SiblingInfo,
@@ -764,7 +765,7 @@ export function createChatController() {
     switch (m.type) {
       case "ready":
         $("nav-model").textContent = m.model || ($("nav-model").textContent as string) || "no model";
-        composer.visionCapable = !!m.vision; updateAttachHint(composer);
+        composer.visionCapable = !!m.vision; composer.transcriptionCapable = !!m.transcription; updateAttachHint(composer);
         composer.thinkingCapable = !!m.thinking; updateThinkingToggle(composer);
         // Per-model recommended sampling (generation_config.json, server-CLI
         // overrides applied) — replaces the old hardcoded SAMP_REC shape.
@@ -979,6 +980,8 @@ export function createChatController() {
       initMemoryPanel();
       // Attachments: + button, file picker, clipboard paste, drag-and-drop.
       ($("chat-attach-btn") as HTMLButtonElement).onclick = () => ($("chat-file-input") as HTMLInputElement).click();
+      // Hold-to-talk mic → /v1/audio/sessions → text into the composer (voice.ts).
+      initVoiceInput($("chat-mic-btn") as HTMLButtonElement, box as HTMLTextAreaElement);
       $("chat-file-input").addEventListener("change", (e) => { addFiles(composer, [...(e.target as HTMLInputElement).files || []]); (e.target as HTMLInputElement).value = ""; });
       box.addEventListener("paste", (e) => {
         const items = e.clipboardData && e.clipboardData.items;

package/src/web/src/composer.ts

@@ -55,6 +55,7 @@ function readText(file: File): Promise<string> {
 export class ComposerState {
   attachments: Attachment[] = [];
   visionCapable = false;  // from the `ready` frame; gates image attachments
+  transcriptionCapable = false;  // from the `ready` frame; shows the hold-to-talk mic
   thinkingCapable = false; // from `ready`; gates the reasoning toggle
   // default on: with tools (read/web_search) the reasoning step gives the
   // small model far better tool judgment (answers math/writing itself,
@@ -179,6 +180,8 @@ export function buildMessageText(state: ComposerState, userText: string): BuiltM
 export function updateAttachHint(state: ComposerState): void {
   const btn = $("chat-attach-btn");
   if (btn) btn.title = state.visionCapable ? "Attach files or images" : "Attach files (this model can't see images)";
+  const mic = $("chat-mic-btn");
+  if (mic) mic.style.display = state.transcriptionCapable ? "" : "none";
 }
 
 /* ────────────────────────────────────────────────────────────────────