gse brain engine: MiMo-V2.6 4-bit on ZeroGPU (free tier)

README.md

@@ -1,13 +1,41 @@
 ---
-title: Mimo Brain Engine
-emoji: 🐨
-colorFrom: indigo
-colorTo: pink
+title: GSE Brain Engine (MiMo-V2.6, ZeroGPU)
+emoji: 🧠
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.29.0
-python_version: '3.13'
+sdk_version: 4.44.0
 app_file: app.py
 pinned: false
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# GSE Brain Engine — MiMo-V2.6 on ZeroGPU
+
+The GSE NFL intelligence engine's LLM specialist layer: Xiaomi **MiMo-V2.6-Distill-Qwen-9B**
+(MIT license) in 4-bit, served on Hugging Face **ZeroGPU — the free tier**.
+
+## Billing policy (from HF official docs, verified 2026-10-02)
+
+- **This Space runs on ZeroGPU hardware only** (`spaces.GPU`, default `large` size).
+  ZeroGPU Spaces are free to host and use — free accounts host up to 2, PRO up to 10.
+- **Daily GPU quotas:** unauthenticated 2 min · free 5 min · PRO 40 min (highest
+  queue priority). Resets 24h after first GPU use.
+- **Overage draws from the pre-paid credit balance only.** No card is charged
+  unless credits were explicitly purchased or auto-recharge is enabled. With no
+  credits, over-quota requests fail/queue — they do not bill.
+- **Paid Space hardware (t4-small, a10g, …) is what bills usage-based.**
+  This Space is pinned to ZeroGPU; a drift monitor alerts loudly if the
+  hardware ever changes. Never upgrade this Space's hardware without
+  Garrett's explicit approval.
+
+Sources: https://huggingface.co/docs/hub/spaces-zerogpu ·
+https://huggingface.co/docs/hub/billing
+
+## The /chat contract
+
+Send VERIFIED DATA plus a grounding prompt. The model reasons (L3 causal
+chains, L4 adversarial review, L5 synthesis) under hard rules: cite provided
+numbers, label inference as INFERENCE with a breaking condition, output the
+requested JSON schema. The deterministic GSE contract (adversary_review,
+correlated_theses, checklist) still disposes — the model proposes.

app.py

@@ -0,0 +1,85 @@
+"""GSE Brain Engine — MiMo-V2.6 on ZeroGPU (free tier).
+
+Serves XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (MIT) in 4-bit behind @spaces.GPU.
+ZeroGPU hardware only — this Space can never bill the account (see README.md
+billing policy). The /chat endpoint speaks the GSE grounding contract: every
+prompt should include VERIFIED DATA and demand cited numbers + INFERENCE flags.
+"""
+
+import spaces
+import torch
+import gradio as gr
+from transformers import (
+    AutoModelForImageTextToText,
+    AutoProcessor,
+    BitsAndBytesConfig,
+)
+
+MODEL_ID = "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B"
+
+SYSTEM_PROMPT = (
+    "You are a specialist in the GSE NFL intelligence engine. "
+    "You are given VERIFIED DATA in the user message. HARD RULES: "
+    "(1) Every numeric claim MUST cite a value from the VERIFIED DATA. "
+    "(2) Anything inferred beyond the data MUST be labeled INFERENCE with a breaking_condition. "
+    "(3) Output ONLY valid JSON when a JSON schema is requested."
+)
+
+print("Loading processor...", flush=True)
+processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+
+print("Loading model in 4-bit...", flush=True)
+bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16)
+model = AutoModelForImageTextToText.from_pretrained(
+    MODEL_ID,
+    quantization_config=bnb,
+    device_map="auto",
+    torch_dtype=torch.float16,
+    trust_remote_code=True,
+)
+model.eval()
+print("Model ready.", flush=True)
+
+
+@spaces.GPU(duration=120)
+def chat(message: str, history: list):
+    """Chat endpoint (api_name=/chat). Text-only; history as Gradio pairs."""
+    try:
+        messages = [{"role": "system", "content": SYSTEM_PROMPT}]
+        for h_user, h_asst in history or []:
+            messages.append({"role": "user", "content": h_user})
+            messages.append({"role": "assistant", "content": h_asst})
+        messages.append({"role": "user", "content": message})
+
+        text = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=False)
+        inputs = processor(text=[text], return_tensors="pt").to("cuda")
+        input_len = inputs["input_ids"].shape[1]
+
+        with torch.inference_mode():
+            out = model.generate(
+                **inputs,
+                max_new_tokens=1024,
+                temperature=0.2,
+                do_sample=True,
+                pad_token_id=processor.tokenizer.eos_token_id,
+            )
+        response = processor.batch_decode(
+            out[:, input_len:], skip_special_tokens=True)[0]
+        return response.strip()
+    except Exception as exc:  # never crash the Space; return the error loudly
+        return f"[BRAIN-ENGINE ERROR] {type(exc).__name__}: {exc}"
+
+
+demo = gr.ChatInterface(
+    fn=chat,
+    title="GSE Brain Engine (MiMo-V2.6, ZeroGPU)",
+    description=(
+        "The GSE NFL intelligence engine's LLM specialist layer, served free on "
+        "Hugging Face ZeroGPU. Send it VERIFIED DATA + a grounding prompt; it "
+        "reasons (L3 chains, L4 adversary, L5 synthesis) and cites numbers."
+    ),
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+gradio>=4.44.0
+torch>=2.8.0
+transformers>=4.57.0
+accelerate>=1.0.0
+bitsandbytes>=0.44.0
+spaces>=0.28.0
+pillow>=10.0.0