Ornith-1.0-35B-NVFP4

NVFP4 (W4A4) quantization of deepreinforce-ai/Ornith-1.0-35B — DeepReinforce's self-scaffolding agentic-coding model (qwen3_5_moe, 35B MoE with a Qwen3-VL vision tower). Quantized with llm-compressor to compressed-tensors nvfp4-pack-quantized.

21.9 GB (from 70.3 GB bf16). Serves on a pair of 16 GB GPUs. Loads in vLLM with no --quantization flag (auto-detected).

What was quantized

All linear layers → NVFP4 (W4A4, group size 16). Kept in bf16: the vision tower (re:.*visual.*), the MoE routers (mlp.gate, mlp.shared_expert_gate), and lm_head. The 30,720 routed-expert projections (256 experts × 3 × 40 layers) are per-expert pack-quantized.

# recipe.yaml
QuantizationModifier:
  targets: [Linear]
  ignore: [lm_head, 're:.*visual.*', 're:.*mlp.gate$', 're:.*mlp.shared_expert_gate$']
  scheme: NVFP4

Benchmarks

pass@1 on HumanEval+ / MBPP+, scored with an identical local harness. Quantized (this model) vs. a panel of same-class open baselines:

Benchmark	no-think	think
HumanEval+ (N=163)	87.1%	93.9%
MBPP+ (N=160)	78.1%	80.6%

With reasoning enabled, the W4A4 quant matches or tops the strongest same-class open coders we benchmarked against, on both suites. Quality of the W4A4 quantization is intact.

Reasoning-model eval tip: Ornith reasons at length. For one-shot code benchmarks (a) give it room (max_tokens ≥ 6500), and (b) extract the answer from after </think> — a naive code extractor that scans the whole message will grab draft code from inside the reasoning block and badly under-score the model.

Throughput (vLLM, NVFP4, on RTX PRO 2000 Blackwell 16 GB)

Config	single-stream	aggregate @ C=8	aggregate (peak)
TP=2	114 tok/s	466 tok/s	~986 tok/s (saturates @ C=32)
TP=4	166 tok/s	699 tok/s	~2280 tok/s (still scaling @ C=64)

(--enforce-eager costs ~5× single-stream; the numbers above are with CUDA graphs on.)

Serving (vLLM)

This box has no NVLink/P2P, hence the NCCL flags. Drop them on a P2P-capable host.

vllm serve sakamakismile/Ornith-1.0-35B-NVFP4 \
  --tensor-parallel-size 2 \
  --max-model-len 8192 \
  --disable-custom-all-reduce \
  --trust-remote-code
# env: NCCL_P2P_DISABLE=1   (no-NVLink hosts only)

Toggle reasoning per request with chat_template_kwargs: {"enable_thinking": true|false}.

Attribution & License

Base model © DeepReinforce, released under MIT. This quantized derivative is redistributed under the same MIT license. All credit for the model itself goes to the original authors — see their model card and technical write-up. This repository only adds the NVFP4 weights and serving metadata.