Ornith 1.0 35B Abliterated MXFP4 Vision MTPLX
Vision-enabled 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. The package contains model files only: no hosted endpoint, telemetry, prompt logs, or external service requirement.
Notes
- Includes the Ornith vision tower and processor files.
- Optimized for MTPLX MTP serving, not LM Studio indexing.
- Uses a compatible prequantized q5/g64 MTP sidecar; recommended draft depth is
2.
- Chat template defaults to thinking off unless
enable_thinking=true is passed explicitly.
- Abliteration metadata is included for transparency; no source direction file is required for inference.
Local Validation
Hardware reference: Apple M4 Max Apple Silicon with 64 GB unified memory.
| Check | Result |
|---|
| API health | Pass |
| Text JSON smoke | Pass |
| Executable code smoke | 3/3 |
| Vision smoke | Pass |
| Mean decode speed | 158.0 tok/s |
| Accepted draft ratio | 92.3% |
The image smoke identified HUNTER and a living room scene. These are lightweight local checks, not public leaderboard scores.
Recommended MTPLX Settings
Use depth 2, profile sustained, tokenizer chat template, MTP enabled, and thinking disabled by default.
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 | 5/10 | 50.0% | 9.60 s | 160.47 tok/s | 35.28 GiB |
| Long context | 1/15 | 6.7% | 36.35 s | 131.56 tok/s | 37.96 GiB |
| Tool calling | 5/6 | 83.3% | 1.32 s | 131.81 tok/s | 36.45 GiB |
| Agentic | 1/2 | 50.0% | 3.30 s | — tok/s | — |
| Vision | 0/4 | 0.0% | 1.72 s | 165.93 tok/s | 35.15 GiB |
- Tested model revision:
afed7f14a357c9db38bf67eb5b7c66b6152157f5
- Benchmark: Shiftedx Bench v0.3.0
- Context lengths represented: 4,096, 16,384, 65,536, 131,072 prompt tokens; effective tested context: not established at the benchmark's 90% threshold
- Runtime contract: MTPLX 2.7.1 sustained; thinking on/medium; sampler temperature=0.2, top_p=0.95, top_k=20; Hermes hybrid routing with native explicit-parallel calls; KV cache
off; MTP depth 2
- Host: Apple M4 Max, 64 GiB unified memory
- Total measured request wall time: 662.69 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.
