Views
No views yet
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.step3p7_mlx.py bridge, which loads the nested Step3p5 text model using MLX and drops vision tensors for text generation.model.safetensors.index.jsonweight_scale, weight_scale_2, or input_scale sidecars are present in the output indexjang_config.json capability verification passes1{
2 "prompt": "What is 2+2? Answer with only the number.",
3 "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",
4 "prompt_tokens": 26,
5 "generated_tokens": 58,
6 "prefill_s": 9.161997079849243,
7 "contains_final_4": true
8}1{
2 "prefill_s": 9.369971990585327,
3 "warm_tokens": 4,
4 "measured_tokens": 32,
5 "decode_s": 0.7534263134002686,
6 "tok_s": 42.47263392697507
7}JANG_2Lmx.quantize1282, 3, 4, 6, 882Gstep3p7_mlx.py wraps mlx_lm.models.step3p5 for text-only proofself_attn.{q,k,v,o,g}_proj: 8-bitembed_tokens: 6-bitgate_proj=4, down_proj=3, up_proj=2jang_config.json stamps:1{
2 "reasoning_parser": "qwen3",
3 "tool_parser": "step3p5",
4 "think_in_template": true,
5 "supports_tools": true,
6 "supports_thinking": true,
7 "family": "step3p7",
8 "modality": "vision",
9 "cache_type": "kv"
10}<think>. Runtimes should not add a second synthetic reasoning prefix.vit_large_projector. No audio tensors or audio tokenizer files were present in the downloaded checkpoint.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.