Shiftedx/ornith-1.0-35b-mxfp8-mlx

Ornith-1.0-35B MLX MXFP8 Vision

This is an unofficial community MLX MXFP8 quantization of deepreinforce-ai/Ornith-1.0-35B, prepared by shiftedx for Apple Silicon and LM Studio.

The build is vision-enabled. It combines the quantized language model with an MLX-compatible vision_tower shard and processor files.

Build

• Base model: deepreinforce-ai/Ornith-1.0-35B
• Format: MLX safetensors
• Quantization: MXFP8, 8-bit, group size 32
• MoE router/gate layers: 8-bit affine
• Indexed tensor bytes: 36,638,678,752
• Indexed parameters: 35,107,181,936
• Shards: 28 safetensors files
• Vision tensors: 333
• Context metadata: 262,144 max context

Compatibility

Validated locally in LM Studio on Apple Silicon:

• Load key: ornith-1.0-35b-mxfp8-mlx
• Runtime context used for validation: 32,768
• Resident memory at 32k context: about 34.15 GiB
• Text smoke: passed
• Vision smoke: passed

The vision config includes a small compatibility adjustment for current MLX/LM Studio loaders: vision_config.model_type is set to qwen3_5_moe.

Lightweight Validation

This is not an official HumanEval leaderboard result. It is a deterministic local smoke test intended to catch obvious quantization or packaging regressions.

Test	Result
HumanEval test[:20] via LM Studio /v1/completions	18/20
Pass rate	90%
Temperature	0
Max tokens	512
Harness note	First-line indentation normalization enabled

Failures in the local smoke run:

• HumanEval/8
• HumanEval/19

Suggested Generation Settings

For general use:

• Temperature: 0.6
• Top-p: 0.95
• Top-k: 20
• Min-p: 0

For deterministic code evaluation, use temperature 0.

Attribution

This quantization is derived from deepreinforce-ai/Ornith-1.0-35B. The upstream model card declares the model MIT licensed. This repository is not an official Deep Reinforce release.