stancsz/Wrench-4B-Qwen3.6-8E-NVFP4-native4M-Experimental-Preview

[!WARNING] Experimental Preview. Not for production use. This model is published for research and preview evaluation only. Do not download it for production deployment or safety-critical workflows. Do not treat benchmark results as production validation. Behavior, APIs, weights, and packaging may change without notice. Use only in an isolated, non-production environment.

Wrench-4B-Qwen3.6-8E Experimental Preview

Wrench is a bounded developer-tool execution SLM for fast, repetitive, verifiable mechanical work. It proposes structured actions or abstains. An independent verifier and the stronger-model fallback retain final authority. Wrench never executes arbitrary shell commands, uses credentials, or writes autonomously.

This experimental candidate is derived from Qwen3.6-35B-A3B, uses NVFP4 W4A16 weights, and contains 3,881,244,016 verified parameters, below the 4.25B parameter ceiling. It is not a general coding agent.

Copy the package

hf download stancsz/Wrench-4B-Qwen3.6-8E-NVFP4-native4M-Experimental-Preview `
  --local-dir Wrench-4B-Qwen3.6-8E-NVFP4-native4M-Experimental-Preview
Set-Location Wrench-4B-Qwen3.6-8E-NVFP4-native4M-Experimental-Preview
.\run_wrench.ps1

The package is a Hugging Face Safetensors directory with tokenizer, bundled verifier, deterministic toolbelt, context policy, and a model-local server. It accepts Ollama-shaped /api/chat and /api/generate requests, including options.num_ctx=4000000, without an external gateway.

What the 4M claim means

The production-value path is hybrid and model-local:

1. 1.receive the complete raw payload directly at the package endpoint;
2. 2.use deterministic MapReduce, bounded search, AST/dependency extraction, and exact lookup windows to identify useful evidence;
3. 3.keep recent intent hot and old material reference-only;
4. 4.compact model work to a bounded effective context, normally 64K;
5. 5.run the bounded proposal, verifier, and identical stronger-model fallback.

The first-layer receipt records selected and omitted spans, raw payload hash, effective working context, route source, and gate latency. Dense native 2M/4M attention is optional research. It is not the Wrench product claim.

When the optional dense-native lane is enabled, -NativeDirectInput enables the first-layer gate automatically. -DenseNativeGate can also be supplied explicitly. The raw request still arrives at the model-local native endpoint, but the package's first model-side stage compacts it to 32K to 64K before expensive attention:

.\serve_freetoken.ps1 -OllamaApi -NativeDirectInput -DenseNativeGate

The gate is fail-closed and records the raw payload hash, selected working context, and gate stage. -BypassDenseNativeGate is reserved for separate reducer-bypassed native capacity probes and cannot be combined with a dense-native mode switch.

Run the model-local endpoint

python .\wrench_server.py --model-dir . --allowed-root . --mechanical-only

The endpoint is http://127.0.0.1:28900/v1/chat/completions. The same process also exposes /api/tags, /api/show, /api/chat, and /api/generate for clients that expect an Ollama-shaped surface. It receives the full raw conversation directly and emits hash-bound context-gate receipts.

The --mechanical-only mode is the verified fast path. Remove it only when a compatible local native backend is available for ambiguous requests. Native generation is separately verified and must not be inferred from the API shape.

The embedded worker API is also available directly from the downloaded directory:

from wrench_worker import WrenchWorker

worker = WrenchWorker.from_pretrained(
    "./Wrench-4B-Qwen3.6-8E-NVFP4-native4M-Experimental-Preview", load_model=False
)
result = worker.propose([
    {"role": "user", "content": "Read README.md with a 65536 byte limit."}
])

Current evidence

On the historical 220-case diagnostic replay, with client-side mechanical shortcut disabled:

• —weighted mechanical frontier-token coverage: 99.2986%;
• —net frontier-token savings: 100%;
• —Wrench plus identical MiniMax fallback final success: 99.6503%;
• —median / p95 latency: 185.095 ms / 299.445 ms;
• —prohibited accepts: 0;
• —unexpected mutations: 0.

The current v102 package-local 4M handoff accepted 3,996,267 raw estimated tokens, staged 1,955 effective model-prefill tokens, and completed in 238.189 ms, including 31.663 ms context-gate time. The upstream in this handoff was a local protocol stub, so this is direct raw intake and bounded MapReduce handoff evidence, not native dense decoder quality.

These are hybrid model-local diagnostics, not dense native attention quality, stock Ollama native generation quality, family-disjoint approval, or production enablement. The current stock Ollama native generation boundary is explicitly recorded as failed on the validation host. GGUF and vLLM require architecture adapters and are not claimed as verified.

Development status

The full source regression is 176 passed. Final release still requires the human-approved family-disjoint MiniMax-worker trace set, independent RTX 5060 Ti verification, and operational shadow evidence. Wrench has no direct mutation authority. It proposes bounded actions or abstains, and the surrounding verifier must enforce execution policy.