ornith-1.5-35b-a3b-affine4-router8-bf16recurrence-vision-mtplx

Affine 4-bit language body with 8-bit MoE router gates and recurrent-state inputs retained in BF16, plus BF16 vision and native BF16 MTP. Built from ornith-ai/Ornith-1.5-35B-A3B revision fbb995a for Apple Silicon.

Format

• Logical size: 24.29 GB
• Qwen3.5 MoE multimodal: 35B total parameters, 3B active
• Language body: affine 4-bit/group-32 with 80 MoE router gates at affine 8-bit/group-64; 60 recurrent input modules retained in BF16
• Vision: 333 same-parent BF16 tensors
• MTP: 785 same-parent BF16 tensors in mtp/weights.safetensors
• Context metadata: 262,144 tokens
• Tokenizer, chat template, and image/video processor metadata preserved

The fail-closed precision recipe and conversion assertions are included in BUILD_RECIPE.json and conversion_receipt.json.

The matched AR/D1/D2/D3 sweep selected D1 at 118.50 decode tok/s on the qualification host.

Usage

mtplx serve \
  --model Shiftedx/ornith-1.5-35b-a3b-affine4-router8-bf16recurrence-vision-mtplx \
  --profile turbo \
  --generation-mode mtp \
  --load-mtp \
  --depth 1

python -m mlx_vlm.generate \
  --model Shiftedx/ornith-1.5-35b-a3b-affine4-router8-bf16recurrence-vision-mtplx \
  --image image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256

Qualification and limitations

• Structural loading, deterministic text, vision, and MTPLX tensor gates passed locally.
• This control quant is experimental; behavior may differ from the BF16 parent and the higher-precision hybrid.
• Full BF16 parent parity was not run on the 64 GiB qualification host.
• Shiftedx Bench results below are tied to the immutable evaluated weight revision.
• Review the upstream model card for intended use, training details, license, and safety considerations.

Shiftedx Bench post-publication qualification

This table was generated from the frozen lightweight quant gate after the model weights were published. Categories remain separate; the benchmark does not produce a composite intelligence score.

Lane	Passed	Accuracy	Mean wall time	Mean decode	Peak active memory
Quality	6/10	60.0%	7.60 s	126.10 tok/s	25.39 GiB
Long context	11/15	73.3%	37.04 s	105.99 tok/s	36.42 GiB
Tool calling	6/6	100.0%	1.54 s	96.90 tok/s	26.62 GiB
Agentic	1/2	50.0%	5.37 s	— tok/s	—
Vision	2/4	50.0%	1.35 s	123.72 tok/s	28.54 GiB

• Tested model revision: bbaff3c9946d1c2fcd30f88138cf8db2970629f1
• Benchmark: Shiftedx Bench v0.3.0
• Context lengths represented: 4,096, 16,384, 65,536, 131,072 prompt tokens; effective tested context: 131,072 tokens
• Runtime contract: MTPLX 2.7.1; MLX 0.32.0; mlx-lm 0.31.3; D1; turbo; thinking on; reasoning effort medium; KV cache off; MTP depth 1
• Host: Apple M4 Max, 64 GiB unified memory
• Total measured request wall time: 656.97 seconds
• 260,096-token status: not run; it is outside the lightweight quant gate.

Scores are specific to the linked model revision, benchmark revision, runtime contract, and host. Changing weight precision, KV-cache precision, reasoning mode, template, or speculative depth creates a different benchmark candidate.