Ornith-1.0-9B is an agentic coding & reasoning model from the DeepReinforce team, post-trained on top of Qwen 3.5 (9B-Dense). It is a reasoning model: each assistant turn opens with a <think> … </think> block before the final answer. Recommended sampling: temperature=0.6, top_p=0.95, top_k=20.

Text tower only. Ornith-1.0-9B is distributed in a Qwen3.5-VL multimodal container (Qwen3_5ForConditionalGeneration), but its trained capabilities are text / code. This MLX conversion keeps the language tower only — the vision encoder is dropped at conversion time by mlx-lm's qwen3_5 loader. Use it as a text/code LLM (it has no image input).

A note on the benchmark

Ornith's headline benchmarks (Terminal-Bench 2.1, SWE-bench Verified/Pro/Multilingual, NL2Repo, ClawEval, SWE Atlas) are agentic harnesses — they need Docker, a large GPU, and long multi-turn rollouts at up to 256K context, so they are not reproducible on a Mac. As an on-device check that quantization preserves coding ability, this card reports HumanEval pass@1 (30 problems, 4096-token budget, with reasoning enabled and <think> stripped before grading). Both MX FP4 and MX FP8 score 93.3% (28/30) with 100% syntactic validity — i.e. 4-bit quantization is effectively lossless on this proxy. HumanEval is a narrow code-generation probe and is not a substitute for the agentic evaluations on the base model card.

This is the reference variant. MX FP8 is near-lossless and is used as the quality/performance reference for the other variants in this collection, so this card reports its absolute numbers (no comparison column).

Benchmark results

Evaluated on Apple M5 Pro with MLX. Model loaded once; performance and quality measured in a single pass.

Performance

	This model	mxfp8 (reference)
Decode tok/s (avg, long traces)	32.14	N/A
Peak memory (GB)	9.672	N/A
Disk size (MB)	8826	8826

Quality

Benchmark	This model	mxfp8 (reference)	n
HumanEval (code, pass@1)	93.3%	N/A	30

Context scaling (decode tok/s)

Context length	Decode tok/s
~128 tokens	32.1
~256 tokens	32.8
~512 tokens	33.3
~1024 tokens	33.4

Usage

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("sahilchachra/ornith-1.0-9b-mxfp8-mlx")
response = generate(model, tokenizer, prompt="Your prompt here", max_tokens=256, verbose=True)

All variants in this collection

Model	Variant
sahilchachra/ornith-1.0-9b-mxfp4-mlx	Block float MX FP4
sahilchachra/ornith-1.0-9b-mxfp8-mlx	Block float MX FP8 ← this model

Notes

Requires Apple Silicon (M1 or later) with MLX
Benchmarks run on Apple M5 Pro, 24 GB unified memory
License: see deepreinforce-ai/Ornith-1.0-9B for the original model's license

Original model

See deepreinforce-ai/Ornith-1.0-9B for full model details and intended use.