add Tab2 live A/B vs Bonsai (sequential: Bonsai first, then POCKET); 2nd llama-server for Bonsai Q1_0; README=BONSAI vs POCKET

Dockerfile

@@ -26,9 +26,12 @@ RUN chmod +x start.sh
 
 ENV MODEL_REPO=FINAL-Bench/POCKET-35B-GGUF \
     MODEL_FILE=POCKET-35B-Q2_K.gguf \
-    N_CTX=4096 \
+    BONSAI_REPO=prism-ml/Bonsai-27B-gguf \
+    BONSAI_FILE=Bonsai-27B-Q1_0.gguf \
+    N_CTX=2048 \
     HF_HOME=/data/hf \
     BACKEND=http://127.0.0.1:8080 \
+    BONSAI_BACKEND=http://127.0.0.1:8081 \
     PORT=7860
 
 EXPOSE 7860

README.md

@@ -1,6 +1,6 @@
 ---
-title: POCKET-35B CPU Chat
-emoji: 📱
+title: POCKET vs Bonsai · CPU
+emoji: ⚔️
 colorFrom: yellow
 colorTo: green
 sdk: docker
@@ -8,43 +8,17 @@ app_port: 7860
 pinned: true
 license: apache-2.0
 models:
-- Qwen/Qwen3.6-35B-A3B
-- Qwen/Qwen3.5-35B-A3B
 - FINAL-Bench/POCKET-35B-GGUF
-- FINAL-Bench/POCKET-KR-GGUF
-- FINAL-Bench/POCKET-EN-GGUF
-- FINAL-Bench/POCKET-KR-MLX
 - FINAL-Bench/Darwin-36B-Opus
-- FINAL-Bench/Darwin-36B-KR
-- FINAL-Bench/Darwin-35B-A3B-Opus
-- FINAL-Bench/Darwin-35B-A3B-Opus-Q8-GGUF
-- FINAL-Bench/Darwin-28B-Opus
-- FINAL-Bench/Darwin-28B-Coder
-- FINAL-Bench/Darwin-28B-Coder-GGUF
-- FINAL-Bench/Darwin-28B-REASON
-- FINAL-Bench/Darwin-28B-KR-Legal
-- FINAL-Bench/Darwin-31B-Opus
-- FINAL-Bench/Darwin-27B-Opus
-- FINAL-Bench/Darwin-60B-DUO
-- FINAL-Bench/Darwin-398B-JGOS
-- FINAL-Bench/Darwin-9B-Opus
-- FINAL-Bench/Darwin-9B-NEG
-- FINAL-Bench/Darwin-9B-MFP4
-- FINAL-Bench/Darwin-4B-Chimera
-- FINAL-Bench/Darwin-4B-David
-- FINAL-Bench/Darwin-4B-Genesis
-- FINAL-Bench/Darwin-TTS-1.7B-Cross
-- FINAL-Bench/Aether-7B-5Attn
-- FINAL-Bench/Aether-7B-5Attn-it
-- FINAL-Bench/Ourbox-35B-JGOS
-- FINAL-Bench/Ourbox-35B-JGOS-GGUF
-- FINAL-Bench/Qwen3.5-35B-A3B-VKAE
-short_description: A 35B model answering on a CPU-only box. No GPU.
+short_description: BONSAI vs POCKET — 35B MoE out-runs the top 1-bit 27B on CPU
 ---
 
-# POCKET-35B · CPU chat
+# BONSAI vs POCKET · live A/B on a CPU
 
-A live demo of **POCKET-35B** (Q2_K, 13 GB) answering on an **upgraded CPU** Space — **no GPU**.
+A live demo of **POCKET-35B** (Q2_K, 13 GB) answering on an **upgraded CPU** Space — **no GPU** —
+with a second tab that races it **head-to-head against Bonsai-27B** (the most-downloaded 1-bit
+on-device model) on the same box, same stock `llama.cpp`. Bonsai answers first, then POCKET —
+watch the tok/s.
 
 POCKET is a sparse Mixture-of-Experts model (34.66B total, ~3B active/token) quantized from
 [Darwin-36B-Opus](https://huggingface.co/FINAL-Bench/Darwin-36B-Opus). It runs on stock

app.py

@@ -1,98 +1,132 @@
 # -*- coding: utf-8 -*-
-"""POCKET-35B CPU chat — FastAPI frontend proxying to a local llama.cpp server.
-Backend is upstream llama.cpp (prebuilt release, supports POCKET's qwen35moe arch).
-Uses the native /completion endpoint with a manually-built ChatML prompt (Qwen-family),
-so it does not depend on the embedded chat template. No GPU."""
+"""POCKET-35B CPU chat + live A/B vs Bonsai.
+FastAPI frontend proxying to two local llama.cpp servers (upstream prebuilt release,
+supports POCKET's qwen35moe arch). Native /completion + manual ChatML. No GPU."""
 import os, json, time
 import httpx
 from fastapi import FastAPI, Request
 from fastapi.responses import HTMLResponse, StreamingResponse, JSONResponse
 
-BACKEND = os.environ.get("BACKEND", "http://127.0.0.1:8080")
+BACKEND = os.environ.get("BACKEND", "http://127.0.0.1:8080")            # POCKET
+BONSAI_BACKEND = os.environ.get("BONSAI_BACKEND", "http://127.0.0.1:8081")  # Bonsai
 HERE = os.path.dirname(__file__)
 
 app = FastAPI(title="POCKET-35B CPU chat")
 
 
 def build_chatml(messages):
-    """Qwen-family ChatML prompt."""
+    """Qwen-family ChatML prompt (works for both POCKET and Bonsai)."""
     parts = []
-    has_system = any(m.get("role") == "system" for m in messages)
-    if not has_system:
-        parts.append("<|im_start|>system\nYou are POCKET, a helpful assistant running on-device.<|im_end|>\n")
+    if not any(m.get("role") == "system" for m in messages):
+        parts.append("<|im_start|>system\nYou are a helpful assistant running on-device.<|im_end|>\n")
     for m in messages:
-        role = m.get("role", "user")
-        content = m.get("content", "")
-        parts.append("<|im_start|>%s\n%s<|im_end|>\n" % (role, content))
+        parts.append("<|im_start|>%s\n%s<|im_end|>\n" % (m.get("role", "user"), m.get("content", "")))
     parts.append("<|im_start|>assistant\n")
     return "".join(parts)
 
 
-@app.get("/api/status")
-async def status():
+def _sse(obj):
+    return "data: " + json.dumps(obj) + "\n\n"
+
+
+async def _health(backend):
     try:
         async with httpx.AsyncClient(timeout=3.0) as c:
-            r = await c.get(BACKEND + "/health")
-        if r.status_code == 200:
-            return JSONResponse({"state": "ready", "detail": "POCKET-35B on CPU · ready"})
-        return JSONResponse({"state": "loading", "detail": "warming up the model…"})
+            r = await c.get(backend + "/health")
+        return r.status_code == 200
     except Exception:
-        return JSONResponse({"state": "loading",
-                             "detail": "downloading & loading POCKET-35B… (first boot can take a few minutes)"})
+        return False
 
 
-@app.post("/api/chat")
-async def chat(req: Request):
-    body = await req.json()
-    messages = body.get("messages") or [{"role": "user", "content": body.get("prompt", "Hi")}]
+async def run_one(backend, tag, messages, max_tokens, temperature=0.7):
+    """Stream one model's /completion output as SSE, tagged with `tag` (None = untagged)."""
     payload = {
         "prompt": build_chatml(messages),
-        "temperature": float(body.get("temperature", 0.7)),
+        "temperature": float(temperature),
         "top_p": 0.95,
-        "n_predict": int(body.get("max_tokens", 512)),
+        "n_predict": int(max_tokens),
         "stream": True,
         "cache_prompt": True,
         "stop": ["<|im_end|>", "<|im_start|>"],
     }
+    n_tok = 0
+    t_first = None
+    tagd = {"m": tag} if tag else {}
+    async with httpx.AsyncClient(timeout=None) as c:
+        async with c.stream("POST", backend + "/completion", json=payload) as r:
+            if r.status_code != 200:
+                detail = (await r.aread()).decode("utf-8", "ignore")[:200]
+                yield _sse(dict(tagd, error="backend %s: %s" % (r.status_code, detail)))
+                return
+            async for line in r.aiter_lines():
+                if not line or not line.startswith("data:"):
+                    continue
+                try:
+                    obj = json.loads(line[5:].strip())
+                except Exception:
+                    continue
+                delta = obj.get("content")
+                if delta:
+                    if t_first is None:
+                        t_first = time.monotonic()
+                    n_tok += 1
+                    yield _sse(dict(tagd, t=delta))
+                if obj.get("stop"):
+                    tim = obj.get("timings") or {}
+                    tok_s = tim.get("predicted_per_second")
+                    if not tok_s and t_first is not None and n_tok > 0:
+                        dt = time.monotonic() - t_first
+                        tok_s = (n_tok / dt) if dt > 0 else 0.0
+                    yield _sse(dict(tagd, done=True, tok_s=round(tok_s or 0.0, 1),
+                                    n=int(tim.get("predicted_n") or n_tok)))
+                    return
+    yield _sse(dict(tagd, done=True, tok_s=0.0, n=n_tok))
+
+
+@app.get("/api/status")
+async def status():
+    pk = await _health(BACKEND)
+    bn = await _health(BONSAI_BACKEND)
+    if pk:
+        detail = "POCKET-35B on CPU · ready" if bn else "POCKET ready · loading Bonsai for the A/B…"
+    else:
+        detail = "downloading & loading models… (first boot can take a few minutes)"
+    return JSONResponse({"state": "ready" if pk else "loading", "detail": detail,
+                         "pocket": pk, "bonsai": bn})
+
+
+@app.post("/api/chat")
+async def chat(req: Request):
+    body = await req.json()
+    messages = body.get("messages") or [{"role": "user", "content": body.get("prompt", "Hi")}]
+    mt = int(body.get("max_tokens", 512))
+
+    async def gen():
+        try:
+            async for chunk in run_one(BACKEND, None, messages, mt, body.get("temperature", 0.7)):
+                yield chunk
+        except Exception as e:
+            yield _sse({"error": str(e)[:200]})
+
+    return StreamingResponse(gen(), media_type="text/event-stream")
+
+
+@app.post("/api/chat_ab")
+async def chat_ab(req: Request):
+    """Sequential race: Bonsai runs first, then POCKET (each gets the full CPU)."""
+    body = await req.json()
+    messages = body.get("messages") or [{"role": "user", "content": body.get("prompt", "Hi")}]
+    mt = int(body.get("max_tokens", 200))
 
     async def gen():
-        n_tok = 0
-        t_first = None
         try:
-            async with httpx.AsyncClient(timeout=None) as c:
-                async with c.stream("POST", BACKEND + "/completion", json=payload) as r:
-                    if r.status_code != 200:
-                        detail = (await r.aread()).decode("utf-8", "ignore")[:200]
-                        yield "data: " + json.dumps({"error": "backend %s: %s" % (r.status_code, detail)}) + "\n\n"
-                        return
-                    async for line in r.aiter_lines():
-                        if not line or not line.startswith("data:"):
-                            continue
-                        try:
-                            obj = json.loads(line[5:].strip())
-                        except Exception:
-                            continue
-                        delta = obj.get("content")
-                        if delta:
-                            if t_first is None:
-                                t_first = time.monotonic()
-                            n_tok += 1
-                            yield "data: " + json.dumps({"t": delta}) + "\n\n"
-                        if obj.get("stop"):
-                            tim = obj.get("timings") or {}
-                            tok_s = tim.get("predicted_per_second")
-                            if not tok_s and t_first is not None and n_tok > 0:
-                                dt = time.monotonic() - t_first
-                                tok_s = (n_tok / dt) if dt > 0 else 0.0
-                            yield "data: " + json.dumps({
-                                "done": True,
-                                "tok_s": round(tok_s or 0.0, 1),
-                                "n": int(tim.get("predicted_n") or n_tok),
-                            }) + "\n\n"
-                            return
-            yield "data: " + json.dumps({"done": True, "tok_s": 0.0, "n": n_tok}) + "\n\n"
+            async for chunk in run_one(BONSAI_BACKEND, "bonsai", messages, mt):
+                yield chunk
+            async for chunk in run_one(BACKEND, "pocket", messages, mt):
+                yield chunk
+            yield _sse({"ab_done": True})
         except Exception as e:
-            yield "data: " + json.dumps({"error": str(e)[:200]}) + "\n\n"
+            yield _sse({"error": str(e)[:200]})
 
     return StreamingResponse(gen(), media_type="text/event-stream")
 

index.html

@@ -3,14 +3,14 @@
 <head>
 <meta charset="utf-8"/>
 <meta name="viewport" content="width=device-width, initial-scale=1"/>
-<title>POCKET-35B · CPU chat</title>
+<title>POCKET-35B · CPU chat & A/B vs Bonsai</title>
 <style>
   :root{ --gold:#f0992a; --teal:#0f9d6e; --ink:#12203a; --muted:#5b6b8f;
-         --line:#e7ecf5; --bg:#f7f9fc; --card:#ffffff; }
+         --line:#e7ecf5; --bg:#f7f9fc; --card:#ffffff; --bonsai:#7c8aa5; }
   *{ box-sizing:border-box; }
   body{ margin:0; font-family:'Segoe UI',Helvetica,Arial,sans-serif; color:var(--ink);
         background:var(--bg); line-height:1.55; }
-  .wrap{ max-width:920px; margin:0 auto; padding:24px 18px 60px; }
+  .wrap{ max-width:960px; margin:0 auto; padding:24px 18px 60px; }
   header{ text-align:center; padding:22px 0 8px; }
   .brand{ font-size:40px; font-weight:800; letter-spacing:1px;
           background:linear-gradient(90deg,#ffd76a,#f0992a); -webkit-background-clip:text;
@@ -21,6 +21,10 @@
   .chip{ background:#fff; border:1px solid var(--line); border-radius:20px; padding:6px 14px;
          font-size:13px; font-weight:600; color:var(--ink); }
   .chip b{ color:var(--teal); }
+  .tabs{ display:flex; gap:8px; justify-content:center; margin:16px 0 4px; }
+  .tab{ background:#fff; border:1px solid var(--line); border-radius:12px; padding:9px 18px;
+        font-size:14.5px; font-weight:700; color:var(--muted); cursor:pointer; }
+  .tab.active{ color:#3a2600; background:linear-gradient(90deg,#ffe6b0,#f7c36a); border-color:#f0992a; }
   .status{ text-align:center; font-size:13px; color:var(--muted); margin:10px 0; }
   .dot{ display:inline-block; width:8px; height:8px; border-radius:50%; background:#c9c8c2;
         margin-right:6px; vertical-align:1px; }
@@ -37,12 +41,12 @@
   .meta b{ color:var(--teal); font-size:13.5px; }
   .empty{ color:var(--muted); text-align:center; padding:40px 10px; font-size:14px; }
   .bar{ display:flex; gap:8px; padding:12px; border-top:1px solid var(--line); }
-  #q{ flex:1; border:1px solid var(--line); border-radius:12px; padding:12px 14px; font-size:15px;
+  #q,#abq{ flex:1; border:1px solid var(--line); border-radius:12px; padding:12px 14px; font-size:15px;
       font-family:inherit; outline:none; }
-  #q:focus{ border-color:var(--gold); }
-  #send{ background:var(--gold); color:#3a2600; border:none; border-radius:12px; padding:0 20px;
+  #q:focus,#abq:focus{ border-color:var(--gold); }
+  #send,#absend{ background:var(--gold); color:#3a2600; border:none; border-radius:12px; padding:0 20px;
          font-weight:700; font-size:15px; cursor:pointer; }
-  #send:disabled{ background:#e7ecf5; color:#9aa6bd; cursor:not-allowed; }
+  #send:disabled,#absend:disabled{ background:#e7ecf5; color:#9aa6bd; cursor:not-allowed; }
   .about{ margin-top:24px; padding:20px 22px; font-size:14.5px; color:#2b3a56; }
   .about h3{ margin:0 0 8px; font-size:16px; }
   .about table{ width:100%; border-collapse:collapse; margin:10px 0; font-size:13.5px; }
@@ -54,6 +58,22 @@
   .ex{ display:flex; gap:8px; flex-wrap:wrap; padding:0 12px 12px; }
   .ex button{ background:#fff; border:1px solid var(--line); border-radius:10px; padding:7px 12px;
               font-size:13px; color:var(--muted); cursor:pointer; }
+  /* A/B */
+  .abhead{ display:grid; grid-template-columns:1fr 1fr; gap:12px; padding:14px 14px 4px; }
+  .abcol{ font-size:15px; font-weight:800; }
+  .abcol span{ display:block; font-size:12px; font-weight:600; color:var(--muted); margin-top:2px; }
+  .abcol.b{ color:var(--bonsai); } .abcol.p{ color:var(--gold); }
+  .abgrid{ display:grid; grid-template-columns:1fr 1fr; gap:12px; padding:6px 14px 4px; }
+  .abpane{ border:1px solid var(--line); border-radius:12px; padding:12px; min-height:200px;
+           max-height:42vh; overflow-y:auto; background:#fbfcfe; }
+  .abpane .txt{ font-size:14px; white-space:pre-wrap; color:var(--ink); }
+  .abpane .m{ font-size:12.5px; font-weight:700; color:var(--muted); margin-top:8px; }
+  .abpane .m b{ font-size:14px; }
+  .abpane.b .m b{ color:var(--bonsai); } .abpane.p .m b{ color:var(--teal); }
+  .abresult{ text-align:center; font-size:17px; font-weight:800; color:var(--ink);
+             padding:12px; min-height:24px; }
+  .abresult b{ color:var(--teal); }
+  @media (max-width:640px){ .abhead,.abgrid{ grid-template-columns:1fr; } }
 </style>
 </head>
 <body>
@@ -70,69 +90,125 @@
     </div>
   </header>
 
+  <div class="tabs">
+    <button class="tab active" data-tab="chat">💬 Chat</button>
+    <button class="tab" data-tab="ab">⚔️ POCKET vs Bonsai</button>
+  </div>
+
   <div class="status"><span id="dot" class="dot loading"></span><span id="st">starting…</span></div>
 
-  <div class="card">
-    <div class="chat" id="chat"><div class="empty" id="empty">Ask anything. This runs on CPU — first reply may take a moment.</div></div>
-    <div class="ex" id="ex">
-      <button>Explain mixture-of-experts in one line.</button>
-      <button>Write a haiku about running AI with no GPU.</button>
-      <button>한국어로 자기소개를 세 문장으로 해줘.</button>
+  <!-- ===== Tab 1: Chat (unchanged) ===== -->
+  <div id="panel-chat">
+    <div class="card">
+      <div class="chat" id="chat"><div class="empty" id="empty">Ask anything. This runs on CPU — first reply may take a moment.</div></div>
+      <div class="ex" id="ex">
+        <button>Explain mixture-of-experts in one line.</button>
+        <button>Write a haiku about running AI with no GPU.</button>
+        <button>한국어로 자기소개를 세 문장으로 해줘.</button>
+      </div>
+      <div class="bar">
+        <input id="q" placeholder="Type a message…" autocomplete="off"/>
+        <button id="send" disabled>Send</button>
+      </div>
     </div>
-    <div class="bar">
-      <input id="q" placeholder="Type a message…" autocomplete="off"/>
-      <button id="send" disabled>Send</button>
+
+    <div class="card about">
+      <h3>What is POCKET?</h3>
+      <p>POCKET is a family of on-device builds quantized from <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus">Darwin-36B-Opus</a>, VIDRAFT's flagship. A 34.66B sparse-MoE model that wakes only ~3B parameters per token — so it runs on a phone, and on a PC with <b>no GPU</b>, using the tools you already have. On the same machine and the same stock <code>llama.cpp</code>, it generates <b>2.7× faster on CPU</b> and <b>2.2× faster on GPU</b> than the leading 1-bit 27B, at matched quality.</p>
+      <table>
+        <tr><th>Build</th><th>Size</th><th>Runs on</th><th>Best for</th></tr>
+        <tr><td>POCKET-35B Q4_K_M</td><td>21 GB</td><td>PC 32 GB</td><td>top quality</td></tr>
+        <tr><td>POCKET-35B <b>Q2_K</b> ⭐ (this demo)</td><td>13 GB</td><td>mini-PC, no GPU</td><td>daily driver</td></tr>
+        <tr><td>POCKET-KR / EN</td><td>~5 GB</td><td>iPhone / Android</td><td>on your phone</td></tr>
+      </table>
+      <div class="links">
+        <a href="https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6">📚 POCKET collection</a>
+        <a href="https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF">POCKET-35B-GGUF</a>
+        <a href="https://huggingface.co/FINAL-Bench/POCKET-KR-MLX">POCKET-KR (iPhone)</a>
+        <a href="https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF">POCKET-EN</a>
+        <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus">🧬 Darwin-36B-Opus</a>
+      </div>
+      <div class="note">This Space runs the Q2_K build on an upgraded CPU (no GPU). Speed here reflects a shared CPU box — a dedicated mini-PC is faster. Apache-2.0 · a VIDRAFT model family.</div>
     </div>
   </div>
 
-  <div class="card about">
-    <h3>What is POCKET?</h3>
-    <p>POCKET is a family of on-device builds quantized from <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus">Darwin-36B-Opus</a>, VIDRAFT's flagship. A 34.66B sparse-MoE model that wakes only ~3B parameters per token — so it runs on a phone, and on a PC with <b>no GPU</b>, using the tools you already have. On the same machine and the same stock <code>llama.cpp</code>, it generates <b>2.7× faster on CPU</b> and <b>2.2× faster on GPU</b> than the leading 1-bit 27B, at matched quality.</p>
-    <table>
-      <tr><th>Build</th><th>Size</th><th>Runs on</th><th>Best for</th></tr>
-      <tr><td>POCKET-35B Q4_K_M</td><td>21 GB</td><td>PC 32 GB</td><td>top quality</td></tr>
-      <tr><td>POCKET-35B <b>Q2_K</b> ⭐ (this demo)</td><td>13 GB</td><td>mini-PC, no GPU</td><td>daily driver</td></tr>
-      <tr><td>POCKET-KR / EN</td><td>~5 GB</td><td>iPhone / Android</td><td>on your phone</td></tr>
-    </table>
-    <div class="links">
-      <a href="https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6">📚 POCKET collection</a>
-      <a href="https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF">POCKET-35B-GGUF</a>
-      <a href="https://huggingface.co/FINAL-Bench/POCKET-KR-MLX">POCKET-KR (iPhone)</a>
-      <a href="https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF">POCKET-EN</a>
-      <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus">🧬 Darwin-36B-Opus</a>
+  <!-- ===== Tab 2: POCKET vs Bonsai (A/B) ===== -->
+  <div id="panel-ab" style="display:none">
+    <div class="card">
+      <div class="abhead">
+        <div class="abcol b">🌳 Bonsai-27B <span>1-bit dense · runs first</span></div>
+        <div class="abcol p">📦 POCKET-35B <span>sparse MoE · runs second</span></div>
+      </div>
+      <div class="abgrid">
+        <div class="abpane b"><div class="txt" id="ab-b-txt"></div><div class="m" id="ab-b-m"></div></div>
+        <div class="abpane p"><div class="txt" id="ab-p-txt"></div><div class="m" id="ab-p-m"></div></div>
+      </div>
+      <div class="abresult" id="ab-result">Type a prompt — Bonsai answers first, then POCKET on the same CPU.</div>
+      <div class="ex" id="abex">
+        <button>Explain how a rocket reaches orbit.</button>
+        <button>Give me 3 tips for better sleep.</button>
+        <button>Write a short poem about the sea.</button>
+      </div>
+      <div class="bar">
+        <input id="abq" placeholder="One prompt — both models answer…" autocomplete="off"/>
+        <button id="absend" disabled>Race</button>
+      </div>
+    </div>
+
+    <div class="card about">
+      <h3>What am I looking at?</h3>
+      <p>The exact same prompt is sent to <b>Bonsai-27B</b> (the most-downloaded on-device model, a 1-bit dense build) and to <b>POCKET-35B</b> (our sparse Mixture-of-Experts, Q2_K) — on the <b>same CPU-only box, the same stock <code>llama.cpp</code></b>. Bonsai runs first, then POCKET. Watch the <b>tok/s</b> on each side.</p>
+      <p>Why POCKET wins on generation: a dense 27B reads all of its weights every token (~3.5 GB); POCKET's MoE wakes only ~3B parameters (~0.7 GB). The weaker the hardware, the wider the gap.</p>
+      <div class="note"><b>Honest notes:</b> sequential race, each model gets the full CPU. tok/s is llama.cpp's measured generation speed. On a shared box a dedicated machine is faster. Prompt <i>processing</i> (not shown) is one place Bonsai leads on very long inputs. Quality is a statistical tie. Bonsai is a trademark of its owners; shown here for factual comparison only.</div>
+      <div class="links">
+        <a href="https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF">📦 POCKET-35B-GGUF</a>
+        <a href="https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6">📚 POCKET collection</a>
+        <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus">🧬 Darwin-36B-Opus</a>
+      </div>
     </div>
-    <div class="note">This Space runs the Q2_K build on an upgraded CPU (no GPU). Speed here reflects a shared CPU box — a dedicated mini-PC is faster. Apache-2.0 · a VIDRAFT model family.</div>
   </div>
 </div>
 
 <script>
-const chat=document.getElementById('chat'), empty=document.getElementById('empty');
-const q=document.getElementById('q'), send=document.getElementById('send');
 const dot=document.getElementById('dot'), st=document.getElementById('st');
-let history=[], busy=false;
+let ready={pocket:false,bonsai:false};
+
+/* ---- tabs ---- */
+document.querySelector('.tabs').addEventListener('click',e=>{
+  if(e.target.tagName!=='BUTTON') return;
+  document.querySelectorAll('.tab').forEach(t=>t.classList.remove('active'));
+  e.target.classList.add('active');
+  const t=e.target.dataset.tab;
+  document.getElementById('panel-chat').style.display = t==='chat'?'':'none';
+  document.getElementById('panel-ab').style.display   = t==='ab'?'':'none';
+});
 
+/* ---- status poll ---- */
 async function poll(){
   try{ const r=await fetch('/api/status'); const s=await r.json();
     st.textContent=s.detail; dot.className='dot '+(s.state==='ready'?'ready':'loading');
-    send.disabled = s.state!=='ready' || busy;
-    if(s.state!=='ready') setTimeout(poll,2500);
+    ready.pocket=!!s.pocket; ready.bonsai=!!s.bonsai;
+    send.disabled = !ready.pocket || busy;
+    absend.disabled = !(ready.pocket&&ready.bonsai) || abBusy;
+    if(!(ready.pocket&&ready.bonsai)) setTimeout(poll,2500);
   }catch(e){ setTimeout(poll,3000); }
 }
-poll();
 
+/* ---- Tab 1 chat ---- */
+const chat=document.getElementById('chat'), empty=document.getElementById('empty');
+const q=document.getElementById('q'), send=document.getElementById('send');
+let history=[], busy=false;
 function add(role,text){
-  if(empty) empty.remove();
+  if(empty&&empty.parentNode) empty.remove();
   const m=document.createElement('div'); m.className='msg '+(role==='user'?'u':'a');
   const b=document.createElement('div'); b.className='bub'; b.textContent=text; m.appendChild(b);
   chat.appendChild(m); chat.scrollTop=chat.scrollHeight; return b;
 }
-
 async function ask(text){
   if(busy||!text.trim()) return; busy=true; send.disabled=true; q.value='';
   add('user',text); history.push({role:'user',content:text});
   const bub=add('assistant','');
-  const meta=document.createElement('div'); meta.className='meta';
-  meta.textContent='⏳ generating on CPU…'; chat.appendChild(meta);
+  const meta=document.createElement('div'); meta.className='meta'; meta.textContent='⏳ generating on CPU…'; chat.appendChild(meta);
   let acc='', nTok=0, t0=0;
   try{
     const r=await fetch('/api/chat',{method:'POST',headers:{'Content-Type':'application/json'},
@@ -154,10 +230,51 @@ async function ask(text){
   }catch(e){ bub.textContent='⚠️ '+e; }
   busy=false; send.disabled=false; q.focus();
 }
-
 send.onclick=()=>ask(q.value);
 q.addEventListener('keydown',e=>{ if(e.key==='Enter') ask(q.value); });
 document.getElementById('ex').addEventListener('click',e=>{ if(e.target.tagName==='BUTTON') ask(e.target.textContent); });
+
+/* ---- Tab 2 A/B race ---- */
+const absend=document.getElementById('absend'), abq=document.getElementById('abq');
+const abTxt={bonsai:document.getElementById('ab-b-txt'),pocket:document.getElementById('ab-p-txt')};
+const abM={bonsai:document.getElementById('ab-b-m'),pocket:document.getElementById('ab-p-m')};
+const abResult=document.getElementById('ab-result');
+let abBusy=false;
+async function race(text){
+  if(abBusy||!text.trim()) return; abBusy=true; absend.disabled=true; abq.value='';
+  abTxt.bonsai.textContent=''; abTxt.pocket.textContent='';
+  abM.bonsai.innerHTML='⏳ Bonsai generating…'; abM.pocket.innerHTML='⏳ waiting for Bonsai to finish…';
+  abResult.textContent='Racing on one CPU…';
+  const acc={bonsai:'',pocket:''}, cnt={bonsai:0,pocket:0}, t0={bonsai:0,pocket:0}, spd={bonsai:0,pocket:0};
+  try{
+    const r=await fetch('/api/chat_ab',{method:'POST',headers:{'Content-Type':'application/json'},
+      body:JSON.stringify({messages:[{role:'user',content:text}],max_tokens:200})});
+    const rd=r.body.getReader(); const dec=new TextDecoder(); let buf='';
+    while(true){ const {done,value}=await rd.read(); if(done) break;
+      buf+=dec.decode(value,{stream:true}); const parts=buf.split('\n\n'); buf=parts.pop();
+      for(const p of parts){ if(!p.startsWith('data:')) continue;
+        const j=JSON.parse(p.slice(5)); const m=j.m;
+        if(j.t && m){ if(!t0[m]) t0[m]=performance.now(); cnt[m]++; acc[m]+=j.t; abTxt[m].textContent=acc[m];
+          if(cnt[m]>1){ const s=cnt[m]/((performance.now()-t0[m])/1000); abM[m].innerHTML='⚡ <b>'+s.toFixed(1)+'</b> tok/s'; }
+          abTxt[m].parentNode.scrollTop=abTxt[m].parentNode.scrollHeight; }
+        if(j.done && m){ spd[m]=j.tok_s||0;
+          abM[m].innerHTML='⚡ <b>'+(spd[m]?spd[m].toFixed(1):'—')+'</b> tok/s · '+(j.n||cnt[m])+' tokens';
+          if(m==='bonsai') abM.pocket.innerHTML='⏳ POCKET generating…'; }
+        if(j.error && m){ abTxt[m].textContent='⚠️ '+j.error; }
+        if(j.ab_done){
+          if(spd.bonsai>0 && spd.pocket>0){ const x=spd.pocket/spd.bonsai;
+            abResult.innerHTML='⚡ POCKET was <b>'+x.toFixed(1)+'× faster</b> than Bonsai — same CPU, no GPU';
+          } else { abResult.textContent='Done.'; } }
+      }
+    }
+  }catch(e){ abResult.textContent='⚠️ '+e; }
+  abBusy=false; absend.disabled=false; abq.focus();
+}
+absend.onclick=()=>race(abq.value);
+abq.addEventListener('keydown',e=>{ if(e.key==='Enter') race(abq.value); });
+document.getElementById('abex').addEventListener('click',e=>{ if(e.target.tagName==='BUTTON') race(e.target.textContent); });
+
+poll();
 </script>
 </body>
 </html>

start.sh

@@ -1,18 +1,18 @@
 #!/usr/bin/env bash
 set -e
 
-# 1) Bring the FastAPI frontend up first, so the page + description render immediately
-#    (chat stays disabled until the backend reports healthy).
+# 1) Frontend up first so the page + description render immediately.
 uvicorn app:app --host 0.0.0.0 --port "${PORT}" &
 
-# 2) Download the GGUF (13 GB — first cold boot can take a few minutes).
-echo "[start] downloading ${MODEL_FILE} from ${MODEL_REPO} …"
-MODEL_PATH=$(python -c "import os;from huggingface_hub import hf_hub_download;print(hf_hub_download(os.environ['MODEL_REPO'],os.environ['MODEL_FILE']))")
-echo "[start] model ready at ${MODEL_PATH}"
+# 2) POCKET backend (:8080) — powers Tab 1 chat AND the POCKET side of the A/B.
+echo "[start] downloading POCKET ${MODEL_FILE} …"
+PK=$(python -c "import os;from huggingface_hub import hf_hub_download;print(hf_hub_download(os.environ['MODEL_REPO'],os.environ['MODEL_FILE']))")
+echo "[start] POCKET at ${PK}"
+llama-server -m "${PK}" --host 127.0.0.1 --port 8080 -c "${N_CTX}" -t "$(nproc)" --jinja --no-webui &
 
-# 3) Start the inference backend: upstream llama.cpp master (supports qwen35moe),
-#    CPU only, model's own chat template via --jinja.
-exec llama-server -m "${MODEL_PATH}" \
-    --host 127.0.0.1 --port 8080 \
-    -c "${N_CTX}" -t "$(nproc)" \
-    --jinja --no-webui
+# 3) Bonsai backend (:8081) — the Bonsai side of the A/B race. Same stock binary.
+echo "[start] downloading Bonsai ${BONSAI_FILE} …"
+BN=$(python -c "import os;from huggingface_hub import hf_hub_download;print(hf_hub_download(os.environ['BONSAI_REPO'],os.environ['BONSAI_FILE']))")
+echo "[start] Bonsai at ${BN}"
+# exec = keep container alive on this process; sequential A/B means no CPU contention.
+exec llama-server -m "${BN}" --host 127.0.0.1 --port 8081 -c "${N_CTX}" -t "$(nproc)" --jinja --no-webui