---
language:
- en
- zh
license: apache-2.0
base_model: stepfun-ai/Step-3.7-Flash-NVFP4
pipeline_tag: image-text-to-text
library_name: mlx
tags:
- mlx
- jang
- jang-k
- stepfun
- vision-language
---

# Step-3.7-Flash-JANG_K

JANG affine conversion of [stepfun-ai/Step-3.7-Flash-NVFP4](https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4).

This JANG_K variant keeps the proven Step JANG text runtime path and uses the routed expert policy:

```text
gate_proj / up_proj / down_proj = 4 / 2 / 2
```

It is the affine K-lane comparison point for the experimental Step `JANGTQ_2K` work.

## Status

Verified locally:

- 58 safetensors shards
- 2,570 indexed tensors
- no raw NVFP4 `weight_scale`, `weight_scale_2`, or `input_scale` sidecars in the output index
- `jang_config.json` capability verification passes
- text generation proof passes through the bundled `step3p7_mlx.py` bridge

Text proof:

```json
{
  "prompt": "What is 2+2? Answer with only the number.",
  "output": "The user is asking \"What is 2+2? Answer with only the number.\" So the answer is 4. The user wants only the number, so I should just output \"4\".\\n</think>\\n4",
  "prompt_tokens": 26,
  "generated_tokens": 43,
  "contains_final_4": true
}
```

Warmed decode proof:

```json
{
  "measured_tokens": 32,
  "decode_s": 0.8008251190185547,
  "tok_s": 39.95878655656726
}
```

## Format

- Format: JANG affine
- Profile: `JANG_K`
- Routed expert policy: `gate_proj=4`, `up_proj=2`, `down_proj=2`
- Attention, router gates, dense/shared MLP, embeddings, and lm head follow the proven Step JANG_2L runtime policy
- Vision/projector tensors are included as F16 passthrough
- Audio tensors: none in the source checkpoint
- MTP tensors: none in the source checkpoint

## Runtime

The bundled `step3p7_mlx.py` bridge maps the nested Step3p7 text config to MLX's Step3p5 text runtime and drops vision tensors for text-only generation.

Required text runtime behavior:

- load `model_file=step3p7_mlx.py`
- preserve the source chat template; it opens the assistant generation prompt inside `<think>`
- use normal KV cache with Step full/sliding attention behavior from the Step3p5 MLX runtime
- do not add a second synthetic reasoning prefix
- use `PreTrainedTokenizerFast`; the source tokenizer metadata otherwise chooses a Llama tokenizer class that decodes byte-level markers incorrectly

Full image-input VLM coherence is not claimed by this artifact. The vision weights are present, but image patch expansion and projector routing still need a Step3p7 VLM wrapper in the target runtime.

## Korean

이 번들은 Step-3.7-Flash-NVFP4를 JANG_K affine `4/2/2` 전문가 비트 정책으로 변환한 산출물입니다. 텍스트 경로는 로컬 MLX 생성 검증을 통과했습니다. 비전 가중치는 포함되어 있지만 이미지 입력 경로는 별도 런타임 구현과 검증이 필요합니다.