---
license: apache-2.0
base_model: vllm-sr/d3-flash
library_name: mlx
tags:
- mlx
- quantized
- decision-model
- image-text-to-text
---

# D3 Flash Q4 for Fenn / MLX

An MLX 4-bit conversion of [vllm-sr/d3-flash](https://huggingface.co/vllm-sr/d3-flash/tree/6838193a0185c0ecfecab937896059d2c3377d61) for local decision verification on Apple silicon.

This is a decision model, not a generative chat model. It requires the Fenn D3 MLX runtime: stock `mlx_lm.generate`, `mlx_vlm.generate`, or Transformers generation do not implement its readout and attention contract.

## Conversion

- Source revision: `6838193a0185c0ecfecab937896059d2c3377d61`.
- Language projections and embeddings: affine 4-bit, group size 64.
- Vision tower: bfloat16.
- Offset RMS normalization weights and decision readout: float32.
- Full-attention layers are noncausal; DeltaNet recurrence remains causal.
- Last-token pooling, source decision temperature, and code ordering are retained.
- `weights.safetensors` includes the decision readout; no generative language-model head is included.
- `fenn_d3.json` records the upstream provenance and SHA-256 checksums of runtime assets.

The weights and tokenizer/processor files are self-contained; the original full-precision checkpoint is not needed at runtime. Text-only and single-image-plus-text evidence are supported by Fenn's verifier.

## Fenn decisions

Fenn supplies three options in this order: A = MATCH, B = NOT_MATCH, C = UNCERTAIN. Their probabilities are scores conditional on the supplied prompt and options, not calibrated accuracy guarantees. Use the matching runtime and evaluate on your own data before choosing thresholds.

## License and attribution

The original model is by the vLLM Semantic Router Team, based on Qwen3.5. The original Apache-2.0 `LICENSE` and `NOTICE` are preserved. This repository changes the weight representation and packages the assets for Fenn's MLX runtime; it does not claim additional training or improved accuracy.