---
license: mit
base_model: ornith-ai/Ornith-1.0-35B
base_model_relation: quantized
pipeline_tag: image-text-to-text
tags:
- nvfp4
- modelopt
- vllm
- blackwell
- awq
- calibrated
- vision
- multimodal
---

# Ornith-1.0-35B — NVFP4, calibrated, with vision

**NVFP4** quantization of `ornith-ai/Ornith-1.0-35B` (Qwen3.5-MoE architecture, multimodal),
built with NVIDIA ModelOpt and — unlike the existing community NVFP4 — **calibrated on real
data**. **Vision works**: the full vision tower ships at BF16 and is verified below on a real
image test (`awq_clip`, 256 samples: 192 UltraChat-200k conversations + 64 real Python source files).
Serves directly in vLLM on Blackwell (built and verified on a GB10, sm_121).

```bash
vllm serve kingjones777/Ornith-1.0-35B-Vision-NVFP4-Calibrated --max-model-len 8192
```

## Measured — not estimated

A/B against `sakamakismile/Ornith-1.0-35B-NVFP4` (the 378K-download dataless build), run
**sequentially on the same idle GB10, same vLLM build, same flags**:

| | this build | community NVFP4 |
|---|---|---|
| decode median (5 samples, 2 matched warm-ups) | **61.29 tok/s** | 60.07 tok/s |
| correctness (`17×23` · capital of Japan · days in 2024) | 3/3 | 3/3 |
| vision (4-quadrant colour image) | 4/4 | 4/4 |
| 10-prompt greedy probe vs the BF16 base | no divergence | no divergence |
| size | 21,900,812,936 B | 21,901,607,200 B |

**Honest summary: speed parity, identical gate results.** The difference is the recipe — this
build's non-expert Linear layers had their quantization clipping calibrated against real
activations; the community build is dataless (`memoryless_minmax`, verified from its shipped
`recipe.yaml`). Whether calibration shows up in your workload is for your evaluation to decide;
no stronger claim is made here.

## Recipe transparency

- Quantized: all `Linear` layers except the vision tower (27 blocks, kept BF16), MoE router
  gates (kept BF16), and `lm_head` (kept BF16) — same protection set as the community build.
- `embed_tokens` is `nn.Embedding` and is **BF16 in both builds** (NVFP4 recipes target `Linear`
  only — an ignore-list entry for it would be a no-op).
- ⚠️ The 30,720 fused-expert projections are quantized **dataless in both builds** — ModelOpt's
  AWQ path cannot reach the transformers-5.x fused expert container. Calibration here covers the
  310 non-expert Linears.
- An `awq_lite` variant was built and **rejected**: it exports `pre_quant_scale` tensors that
  vLLM silently ignores, which would serve incorrect math while appearing to work.

## Verification

Every gate above was run against the served model over the OpenAI API, answers asserted against
content + reasoning with `finish_reason` recorded. File sizes on this repo are byte-verified
against the build box. Built on NVIDIA GB10 (Blackwell), CUDA 13.0, ModelOpt 0.44, vLLM nightly
cu130.

## Credits

Base model by `ornith-ai`. Quantized with NVIDIA ModelOpt. The community NVFP4 by
`sakamakismile` was the benchmark target and its disclosed recipe made the controlled
comparison possible.