Step-3.7-Flash-JANG_2L

Step-3.7-Flash-JANG_2L

JANG_2L conversion of stepfun-ai/Step-3.7-Flash-NVFP4.

This bundle was built from the public NVFP4 checkpoint. Routed MoE tensors were decoded from ModelOpt NVFP4 (uint8 payload, float8_e4m3fn block scales, fp32 side scales) and then re-quantized into JANG affine weight/scales/biases tensors. BF16 attention, shared expert, dense, vision, and projector tensors were handled according to the JANG plan.

Status

This artifact has a text-only local coherence proof through the bundled step3p7_mlx.py bridge, which loads the nested Step3p5 text model using MLX and drops vision tensors for text generation.

Verified locally:

• 67 safetensors shards
• 2,570 tensors in model.safetensors.index.json
• No missing shard references
• No raw NVFP4 weight_scale, weight_scale_2, or input_scale sidecars are present in the output index
• jang_config.json capability verification passes
• Text generation proof passes on a math prompt

Text proof:

{
  "prompt": "What is 2+2? Answer with only the number.",
  "output": "The user asks \"What is 2+2? Answer with only the number.\" So the answer is 4. The user wants only the number. So we should output \"4\". There's no disallowed content. It's a simple arithmetic. So we comply.\\n</think>\\n4",
  "prompt_tokens": 26,
  "generated_tokens": 58,
  "prefill_s": 9.161997079849243,
  "contains_final_4": true
}

Speed note: short cold measurements include MLX graph/kernel compile and are not representative of steady decode. A no-wrapper warmed decode run over 32 measured tokens produced:

{
  "prefill_s": 9.369971990585327,
  "warm_tokens": 4,
  "measured_tokens": 32,
  "decode_s": 0.7534263134002686,
  "tok_s": 42.47263392697507
}

Still required before full VLM runtime claims:

• Step3p7 VLM wrapper in the target MLX/vMLX runtime
• image patch token expansion and vision projector path

Format

• Format: JANG affine
• Profile: JANG_2L
• Quantization backend: mx.quantize
• Default group size: 128
• Bit widths used: 2, 3, 4, 6, 8
• Vision/projector: BF16 source converted to F16 passthrough for this first artifact
• Output size: about 82G
• Runtime bridge: step3p7_mlx.py wraps mlx_lm.models.step3p5 for text-only proof

Important allocation choices:

• self_attn.{q,k,v,o,g}_proj: 8-bit
• embed_tokens: 6-bit
• routed experts: gate_proj=4, down_proj=3, up_proj=2
• true router/gate tensors: passthrough where present

Runtime Metadata

jang_config.json stamps:

{
  "reasoning_parser": "qwen3",
  "tool_parser": "step3p5",
  "think_in_template": true,
  "supports_tools": true,
  "supports_thinking": true,
  "family": "step3p7",
  "modality": "vision",
  "cache_type": "kv"
}

The source chat template opens the assistant generation prompt inside <think>. Runtimes should not add a second synthetic reasoning prefix.

Vision And Audio

The source checkpoint contains the Step vision encoder and vit_large_projector. No audio tensors or audio tokenizer files were present in the downloaded checkpoint.

The source config mentions next-token prediction layers, but no MTP/nextn tensors were present in the NVFP4 source. This bundle does not synthesize MTP tensors from config fields.

Korean

이 번들은 stepfun-ai/Step-3.7-Flash-NVFP4를 JANG_2L 형식으로 변환한 산출물입니다. 텍스트 경로는 step3p7_mlx.py 브리지를 통해 로컬 생성 검증을 통과했습니다. 비전 가중치는 포함되어 있지만, 이미지 입력 경로는 아직 별도 런타임 구현과 검증이 필요합니다. 오디오 텐서는 원본 체크포인트에 없었습니다.