lokinfey/Qwen3_5_4B_Hmm_ONNX

Qwen3.5-4B-Hmm ONNX

CPU-oriented ONNX conversion of n4ze3m/Qwen3.5-4B-Hmm, built from Qwen3.5-4B-Hmm-Q4_K_M.gguf with Microsoft ONNXRuntime Mobius.

Hmm is an experimental Qwen3.5-4B fine-tune that reproduces the typed probabilistic decision idea associated with Jev-style System One workflows. It is not Jev, is not affiliated with TypeSafe AI, and does not inherit Jev's native architecture, performance, calibration, or type-safety guarantees.

Conversion

mobius build-gguf Qwen3.5-4B-Hmm-Q4_K_M.gguf \
  --output onnx_outputs \
  --ep cpu \
  --dtype f32 \
  --release

The model keeps packed quantized storage where supported. See quantization_report.json for the exact Q4_K_M conversion dispositions.

Files

• model.onnx: ONNX graph.
• model.onnx.data: external packed model weights.
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets pinned from Qwen/Qwen3.5-4B.
• quantization_report.json: Mobius conversion report.
• cpu_test_summary.json: direct ONNX Runtime CPU validation results.
• run_hmm_onnx.py: typed-decision inference example.

Keep model.onnx and model.onnx.data in the same directory.

Usage

pip install numpy onnxruntime transformers
python run_hmm_onnx.py --model-dir .

The runner accepts an optional JSON request:

python run_hmm_onnx.py --model-dir . --request request.json

{
  "state": "Help! My payouts have failed for 3 days. I need the money today.",
  "questions": {
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this message convey urgency?"
    },
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments, invoicing, refunds",
        "technical": "Bugs, outages, integrations",
        "sales": "Pricing, upgrades, new accounts"
      }
    }
  }
}

Validation

The exact uploaded artifact was tested with ONNX Runtime 1.30.0 and CPUExecutionProvider:

• ONNX checker: passed
• Stateful hybrid DeltaNet/KV inference: passed
• Hmm typed-decision test: passed
• Example urgent probability: 0.9745
• Example department: billing (0.7216)

Reproducibility

• Source revision: c27fa3c627dfaced343c6ba9d3a0d00243a3be51
• Source GGUF SHA-256: 5e03cb057049c56b421bd3c506d77fd8e0a77996bb8464148a02cfc3caac5229
• Mobius revision: 6b27a3f08b8b5d08ba9b14b416e3b435942bb0bd
• Tokenizer revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a

Conversion notebook and source code: kinfey/JevONNX.

Limitations

• One question requires one model pass.
• The adapter normalizes first-token A/B/C/... probabilities over the supplied options.
• The source model card recommends against using Hmm as the only safety gate or for important decisions.
• Mobius lists qwen35 graph and quantized import as supported, while representative real-weight runtime evidence remains pending. This artifact was therefore validated directly with ONNX Runtime CPU.

Citation

@misc{nazeem2026hmm,
  author = {Muhammed Nazeem},
  title  = {Hmm: a small open model for typed decisions},
  year   = {2026},
  url    = {https://github.com/n4ze3m/hmm}
}