Shiftedx/ornith-1.5-35b-a3b-attention8-bf16recurrence-vision-mtplx

Ornith 1.5 35B A3B — hybrid MLX + vision + MTPLX

Vision-preserving hybrid MLX quantization of Ornith-1.5-35B-A3B, built from upstream revision fbb995a. It is intended for Apple Silicon and includes the native vision tower and MTP sidecar.

Model format

• 28.12 GB; Qwen3.5 MoE multimodal (35B total, 3B active)
• Language body: 4-bit affine, group size 32
• Attention-sensitive modules: 8-bit affine, group size 64
• Recurrent linear-attention projections, vision tensors, and MTP tensors: BF16
• Context metadata: 262,144 tokens; effective tested context: 65,536 tokens
• Included: tokenizer, chat template, image/video processor metadata, build recipe, and conversion receipt

The evaluated public weights were anonymously verified byte-identical to model revision 08ce2da.

Recommended runtime

MTPLX 2.7.1 selected turbo depth 1 (D1). On the qualification prompt, D1 reached 116.95 decode tok/s versus 63.12 tok/s autoregressive, a 1.85× speedup, with both quality checks passing. Use precise KV cache for the benchmarked configuration.

mtplx serve \
  --model Shiftedx/ornith-1.5-35b-a3b-attention8-bf16recurrence-vision-mtplx \
  --profile turbo \
  --generation-mode mtp \
  --load-mtp \
  --depth 1 \
  --reasoning on \
  --reasoning-effort medium

Shiftedx Harness

A separate three-trial Shiftedx Harness evaluation scored 24/90 (26.7%) direct and 85/90 (94.4%) with the harness. This measures a deployment policy layer, not a change to the model weights.

Vision usage

python -m mlx_vlm.generate \
  --model Shiftedx/ornith-1.5-35b-a3b-attention8-bf16recurrence-vision-mtplx \
  --image image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256

Limitations

• This is an experimental quantized runtime artifact; behavior may differ from the BF16 parent.
• Full BF16 parent parity and the 262,144-token metadata limit were not tested.
• Strict JSON output can include Markdown fences without an additional policy layer.
• Benchmark results are specific to the linked revisions, runtime settings, and host.
• 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	8/10	80.0%	10.16 s	108.50 tok/s	42.45 GiB
Long context	13/15	86.7%	38.92 s	94.28 tok/s	45.12 GiB
Tool calling	6/6	100.0%	1.72 s	84.64 tok/s	43.65 GiB
Agentic	1/2	50.0%	4.65 s	— tok/s	—
Vision	1/4	25.0%	1.45 s	108.00 tok/s	42.96 GiB

• Tested model revision: 08ce2daeea94986b447acd6691b76110856de49e
• Benchmark: Shiftedx Bench v0.3.0
• Context lengths represented: 4,096, 16,384, 65,536, 131,072 prompt tokens; effective tested context: 65,536 tokens
• Runtime contract: MTPLX 2.7.1 turbo D1; thinking on/medium; temperature=1.0, top_p=0.95, top_k=20; native tool prompt; tokenizer chat template; KV cache off; MTP depth 1
• Host: Apple M4 Max, 64 GiB unified memory
• Total measured request wall time: 710.85 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.