GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
stancsz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
stancsz
Model Tree
Input Modalities
Output Modalities
Supported Functionality
powershell
hf download stancsz/Wrench-4B-Qwen3.6-8E-Experimental-Preview --local-dir Wrench-4B-Qwen3.6-8EThe canonical distribution format is Hugging Face Safetensors. The package embeds the tokenizer hook, deterministic mechanical lookup runtime, read-only verifier, long-context overlay, hash-bound package metadata, and the FreeToken launcher. It is intended to feel like one model directory, not a separately installed harness.
The embedded worker API is available directly from the downloaded directory:
python
from wrench_worker import WrenchWorker worker = WrenchWorker.from_pretrained("./Wrench-4B-Qwen3.6-8E", load_model=False)result = worker.propose([ {"role": "user", "content": "Read README.md with a 65536 byte limit."}])With load_model=True, ambiguous requests use the standard Transformers model
and still pass through the same fail-closed verifier. High-confidence
mechanical requests use the embedded route without a model call.
For the standard Hugging Face config and tokenizer path, use Transformers 5.17.0 or newer:
powershell
python -m pip install -U "transformers>=5.17.0" huggingface_hubThe bundled launcher requires a compatible FreeToken build and a CUDA GPU:
powershell
.\serve_freetoken.ps1The package declares a 4M input endpoint and uses an 8K recent SWA window in the experimental native profile. The default fast path mechanically reduces noisy payloads to a 64K effective working context. Native direct input and fast staged input are recorded separately in receipts.
The embedded reducer has also passed a 4M mechanical stress diagnostic: 3,999,951 estimated raw tokens reduced to a 92-token model prefill, with 1.0 target-reference recall, 1.0 current-intent preservation, and 1.0 hash-bound reference rate. The measured cold ingest was 98.713 ms and hot selection was 44.441 ms on the local development machine. This is deterministic toolbelt evidence, not an LLM long-context quality or MiniMax-parity claim.
The base checkpoint config is 2M position-capable. The 4M probe uses a runtime RoPE extension and is not long-context training. Native 4M retrieval quality, throughput under production concurrency, and the full MiniMax matched-workflow acceptance suite remain open measurements.
The standard Transformers path has verified architecture/configuration, tokenization, and full weight loading. Generation quality through plain Transformers is not yet a pass and is intentionally not claimed here. Use the bundled FreeToken profile for the current experimental generation path.
Wrench has no direct mutation authority. It proposes bounded developer-tool actions or abstains; a surrounding verifier must enforce execution policy.
from wrench_worker import WrenchWorker worker = WrenchWorker.from_pretrained("./Wrench-4B-Qwen3.6-8E", load_model=False)result = worker.propose([ {"role": "user", "content": "Read README.md with a 65536 byte limit."}])hf download stancsz/Wrench-4B-Qwen3.6-8E-Experimental-Preview --local-dir Wrench-4B-Qwen3.6-8Epython -m pip install -U "transformers>=5.17.0" huggingface_hub.\serve_freetoken.ps1