Shiftedx/ornith-1.0-35b-mxfp4-mtplx

Ornith 1.0 35B MXFP4 MTPLX

Text-only MXFP4 MLX build of deepreinforce-ai/Ornith-1.0-35B, packaged for MTPLX native-MTP inference on Apple Silicon.

This is intended for local, private inference. After download, prompts and outputs can stay on your machine when served with a local MTPLX endpoint.

Notes

• Text-only: no vision tower is included.
• Optimized for MTPLX, not LM Studio.
• Uses a compatible transplanted MTP sidecar; recommended draft depth is 2.
• Reasoning should be routed separately with the Qwen reasoning parser.

Recommended MTPLX Settings

python -m mtplx.server.openai \
  --model /path/to/ornith-1.0-35b-mxfp4-mtplx \
  --backend-id qwen3_next \
  --generation-mode mtp \
  --load-mtp \
  --depth 2 \
  --profile sustained \
  --chat-template-profile tokenizer \
  --normalize-thinking-tags \
  --reasoning-mode on \
  --enable-thinking \
  --reasoning-parser qwen3 \
  --reasoning-effort high \
  --temperature 0.2 \
  --top-p 0.95 \
  --top-k 20 \
  --no-stats-footer

Local Validation

Hardware reference: Apple M4 Max Apple Silicon with 64 GB unified memory.

On a local Apple Silicon host, this MTPLX profile matched the LM Studio text baseline on a small hard validation suite:

Runtime	Score	Measured speed
LM Studio text baseline	7/10	104 tok/s
MTPLX depth 2	7/10	133 tok/s

This is a lightweight local validation, not a public leaderboard result.

Privacy

This repository contains model files only. It does not include a hosted endpoint, telemetry, or an external service requirement. Use a local server and inspect your client configuration if strict data locality matters.