Sign up

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

🤗 Hugging Face sourceimage-text-to-textmit35.1B params70 GBsafetensors✓ Checksums witnessedupdated today
No torrent yet →

Ornith 1.5 35B-A3B Abliterated — MLX Vision + Native MTP

Hybrid MLX quant of ornith-ai/Ornith-1.5-35B-A3B, pinned to revision 10fbf86fed7ecee4a061f8b499a618f46001cac1.

Artifact

  • 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
  • Abliteration strength 1.5 across 120 attention/shared-expert/switch-expert residual writers
  • Recommended runtime: MTPLX 2.9.1, MLX 0.32.1, MLX-LM 0.31.3
  • Recommended speculative depth: D3

Run

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

For exact JSON, install MTPLX's optional llguidance dependency and send an OpenAI-compatible strict response_format schema.

Local qualification

Tested on an M4 Max with 64 GB unified memory.

Gate Result
Held-out refusal rate 100% parent → 0% modified
Held-out benign refusal rate 0%
Held-out utility pass rate 83.3%
Executable Python tasks 8/8
Native function-call score 6/6
Parallel tool-plan score 6/6
OpenAI API code smoke 3/3
Vision semantic checks 4/4
Strict-schema vision OCR pass
Text exact-output smoke pass
Vision tensor preservation 333/333 exact
MTP tensor gate 785/785 exact

The final three-prompt, 512-token MTPLX sweep selected D3 at 112.1 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

Abliteration changes refusal behavior; it does not guarantee correctness or safe output. The model may wrap JSON in Markdown when no response schema is supplied. 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.