4-bit affine expert body with 8-bit attention, routing, embeddings, LM head, and late-layer projections
60 recurrent tensors retained in BF16
333 native BF16 vision tensors
785 native BF16 MTP tensors in mtp/weights.safetensors
Recommended runtime: MTPLX 2.9.1, MLX 0.32.1, MLX-LM 0.31.3
Recommended speculative depth: D3

Run

bash

mtplx serve \
  --model Shiftedx/ornith-1.5-35b-a3b-attention8-bf16recurrence-vision-mtplx \
  --profile sustained \
  --depth 3 \
  --mtp \
  --reasoning off

Local qualification

Tested on an M4 Max with 64 GB unified memory.

Gate	Result
Executable Python tasks	8/8
Text exact-output smoke	pass
Vision smoke	pass
MTP tensor gate	785/785 exact

The final three-prompt, 512-token MTPLX sweep selected D3 at 109.0 decode tok/s. D1, D2, and D3 all passed the quality gate. Fans remained on automatic, so treat the speed as a host-specific measurement rather than a universal benchmark.

Notes

Vision requires a compatible MLX-VLM or MTPLX runtime. Review generated code and tool calls before execution. The parent model and this derivative use the upstream MIT license; see the parent repository for original training and architecture details.