Ornith 1.0 35B · NVFP4
Fit a 34.7B Qwen3.5-MoE reasoning checkpoint onto a single Blackwell workstation without quantizing every subsystem.
Release notes for Ornith 1.0 35B · NVFP4: the problem, the decision, the artifacts on Hugging Face, and what the evaluation showed. Figures come from the committed snapshot of 2026-09-01.
Problem
Fit a 34.7B Qwen3.5-MoE reasoning checkpoint onto a single Blackwell workstation without quantizing every subsystem.
Decision
Quantized routed experts to NVFP4 with NVIDIA ModelOpt; kept attention QKV, shared experts, and the vision encoder at higher precision.
Artifacts
| Artifact | Format | Released | Downloads · 30 days |
|---|---|---|---|
| Ornith-1.0-35B-NVFP4 (opens in a new tab) | NVFP4 | 2026-06-29 | 3,202 |
Result
About 23 GB versus 69 GB BF16, validated to load and generate with vLLM on an RTX PRO 6000 Blackwell.
Serving check
A minimal single-node vLLM start for the serving build; adjust context length and parallelism to the hardware.
# vLLM, single node; tune --max-model-len and --tensor-parallel-size to the GPU
vllm serve ressl/Ornith-1.0-35B-NVFP4 \
--dtype auto \
--max-model-len 32768Evidence
- Model card: Ornith-1.0-35B-NVFP4 (opens in a new tab)
- Case on ressl.ch: https://ressl.ch/#model-ornith-nvfp4
- Hugging Face snapshot captured 2026-09-01