Ornith-1.0-35B-MLX-3bit-MTP

An Apple-Silicon MLX build of deepreinforce-ai/Ornith-1.0-35B that combines the 3-bit multimodal backbone from mlx-community/Ornith-1.0-35B-3bit with an experimental one-layer MTP head extracted from georgeis55/Ornith-1.0-35B-MLX-oQ8-mtp.

This checkpoint is prepared for Lightning MTP speculative decoding in oMLX. The original vision encoder and VLM configuration are retained.

Technical layout

• Architecture: Qwen3.5 MoE VLM, 40 backbone layers, 256 experts
• Backbone: affine 3-bit MLX quantization, group size 64 (3.662 bits/weight)
• MTP: one full-attention MoE prediction layer at language_model.mtp.*
• MTP linears: affine 8-bit, mixed group sizes 64 and 128
• MTP fusion projection and normalization weights: BF16
• MTP file: model-mtp-oq8.safetensors (905,114,024 bytes)
• MTP SHA-256: 26358039bfc747625e8d476fede0cf619401a2ca5da9a03898d9b0c282f65c40
• Tested runtime: oMLX 0.5.3 on Apple Silicon

The weight index contains all 42 MTP tensor mappings. The configuration declares one next-token prediction layer and includes per-module quantization overrides so the 8-bit MTP tensors are not interpreted using the 3-bit backbone settings.

Download and use with oMLX

hf download wd01216-bit/Ornith-1.0-35B-MLX-3bit-MTP \
  --local-dir ~/.omlx/models/wd01216-bit/Ornith-1.0-35B-MLX-3bit-MTP

In oMLX, open the model settings and enable Lightning MTP. The model identifier will be Ornith-1.0-35B-MLX-3bit-MTP when downloaded into the directory above.

Validation

The final checkpoint was discovered and loaded successfully by oMLX 0.5.3 as a multimodal model with Native MTP active. Two short text smoke tests confirmed that the MTP decode path was used, with observed draft acceptance of 15/16 and 6/6. These tiny tests verify wiring and execution only; they are not representative benchmarks.

Important limitation

The MTP head is a verbatim graft originally sourced from a Qwopus3.6 35B model. It was not trained on Ornith hidden states, and the donor bundle was prepared around an oQ8 backbone rather than this 3-bit backbone. Acceptance rate, output quality, memory use, and speedup are workload-dependent and may be lower than those of a native jointly trained MTP head. Disable Lightning MTP if a workload shows poor acceptance or regressions.

Provenance

• Upstream model: deepreinforce-ai/Ornith-1.0-35B
• Quantized multimodal backbone: mlx-community/Ornith-1.0-35B-3bit
• MTP donor bundle and conversion notes: georgeis55/Ornith-1.0-35B-MLX-oQ8-mtp
• Adapter assembly and oMLX validation: 2026-07-27

License

The upstream Ornith model is released under the MIT license. Refer to the upstream model card and license for intended use, limitations, and attribution requirements.