ornith-1.0-35b-mxfp4-vision-mtplx

<p align="center"><img src="assets/ornith_logo.png" alt="Ornith 1.0" width="600"/></p>

ornith-1.0-35b-mxfp4-vision-mtplx

Vision-preserving MLX conversion of deepreinforce-ai/Ornith-1.0-35B, pinned to revision 5df2ed3f675c7beaa490328cc70bb573b65fb660.

Format

• Language body: MXFP4, 4-bit, group size 32
• MoE gate overrides: affine 8-bit, group size 64
• Vision tower: source BF16, 333 tensors
• Architecture: Qwen3.5 MoE multimodal, 40 language layers, 256 experts
• Tokenizer, chat template, processor metadata, and vision features preserved from the pinned source

Local qualification

• Structural inspection: pass (1,658 indexed tensors; 512 quantized weights; 80 MoE gate overrides)
• Shard index: pass; no stale or missing shards
• Strict MLX-LM and MLX-VLM lazy load: pass
• Cross-variant tokenizer, template, processor, and vision-shard parity: pass
• Vision features: source BF16 tensor payload preserved

MTPLX runtime

This self-contained experimental variant combines the qualified MXFP4 Ornith body and BF16 vision tower with a compatible Qwen3.5-MoE Q5/G64 MTP sidecar. Ornith's upstream 35B checkpoint does not include native MTP tensors. The sidecar is byte-identical to mtp.safetensors from Shiftedx/ornith-1.0-35b-abliterated-mxfp4-vision-mtplx at revision afed7f14a357c9db38bf67eb5b7c66b6152157f5 (SHA-256 d9788ed667167b211cdb178247f4d5ccc2213f349611d7c05486d23f6b364b84).

MTPLX 2.0.2 inspection and tensor-contract gates passed. In the frozen one-prompt performance test, depth 2 achieved 136.94 decode tok/s versus 89.41 tok/s autoregressive (+53.17%); end-to-end throughput improved from 82.89 to 123.84 tok/s (+49.40%). Depth-2 acceptance was 94.51% and 87.78% by draft depth. Quality gates passed, but every completion reached the 256-token ceiling, fans remained on automatic, and the figures are a point estimate rather than a statistical benchmark.

mtplx inspect --require-mtp <local-path-or-repo-id>
mtplx serve --model <local-path-or-repo-id> --generation-mode mtp --mtp-depth 2 --reasoning-mode off

Basic use

python -m mlx_vlm.generate \
  --model <local-path-or-repo-id> \
  --image image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256

Review the upstream model card for intended use, training, license, benchmarks, and deployment considerations. Quantization does not change the upstream model's safety characteristics.