---
library_name: transformers
license: mit
license_link: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/blob/main/LICENSE
pipeline_tag: text-generation
base_model: deepreinforce-ai/Ornith-1.0-35B
base_model_relation: quantized
tags:
- fp8
- fp8-block
- compressed-tensors
- llm-compressor
- qwen3.5
- moe
- code
- reasoning
---

<img width="600px" src="assets/ornith_logo.png">

# Ornith-1.0-35B-FP8-BLOCK

This repository contains a public `FP8_BLOCK` / `compressed-tensors`
quantization of [`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B).

The original model is Ornith-1.0-35B, a Qwen3.5 MoE-family reasoning and
agentic coding model released by DeepReinforce. This repository only changes
the checkpoint representation; it does not introduce additional fine-tuning.

This model is a faithful quant of the original Ornith-1.0-35B and therefore **does not** have an MTP head, however, we have also uploaded a [shisa-ai/Ornith-1.0-35B-FP8-BLOCK-MTP](https://huggingface.co/shisa-ai/Ornith-1.0-35B-FP8-BLOCK-MTP) model that grafts the Qwen 3.6 35B-A3B official MTP head onto the model. This is the best-performing MTP and based on our testing it gives a >20% throughput uplift for coding workloads (tested on RTX PRO 6000 and vLLM v0.23.0).

## Quantization Summary

- Source model: [`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B)
- Quantized model: `shisa-ai/Ornith-1.0-35B-FP8-BLOCK`
- Quantization tool: `llm-compressor` model-free PTQ
- Quantization format: `compressed-tensors`
- Scheme: `FP8_BLOCK`
- Calibration data: none; this is data-free/model-free PTQ
- Weight quantization: static FP8, symmetric, block strategy, `128x128` blocks
- Activation quantization: dynamic FP8, symmetric, group strategy, group size `128`
- Target modules: `Linear`
- `compressed-tensors` metadata version recorded in `config.json`: `0.15.1.a20260406`

The local quantization environment used for this artifact reported:

- `llmcompressor=0.10.1.dev67+ga1cec6fa`
- `compressed-tensors=0.15.1a20260406`
- `transformers=5.5.0`

The definitive machine-readable quantization metadata is in
[`config.json`](config.json).

## Unquantized / Ignored Paths

The quantization run intentionally skipped the following module patterns:

```text
re:.*lm_head$
re:.*embed_tokens$
re:.*visual.*
re:.*mlp\.gate$
re:.*mlp\.shared_expert_gate$
re:.*linear_attn.*
re:^mtp.*
```

Practical implications:

- `lm_head` and token embeddings remain unquantized.
- MoE router/gating paths remain unquantized.
- Qwen3.5 linear-attention/Gated-DeltaNet paths remain unquantized.
- Vision tower weights remain unquantized in this artifact.
- No MTP tensors were present in the source checkpoint used here, so this model
  card does not advertise MTP/speculative decoding support.

## Example Usage

Use a recent runtime that supports both `Qwen3_5MoeForConditionalGeneration`
and `compressed-tensors` `FP8_BLOCK` checkpoints.

### vLLM

```bash
vllm serve shisa-ai/Ornith-1.0-35B-FP8-BLOCK \
  --served-model-name Ornith-1.0-35B-FP8-BLOCK \
  --tensor-parallel-size 8 \
  --host 0.0.0.0 \
  --port 8000 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --trust-remote-code
```

Adjust `--tensor-parallel-size`, `--max-model-len`, and memory settings for your
hardware. Long-context serving still requires substantial KV-cache memory even
though the checkpoint weights are compressed.

### OpenAI-Compatible Client

```python
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="Ornith-1.0-35B-FP8-BLOCK",
    messages=[
        {"role": "user", "content": "Write a Python function is_prime(n). Keep it short."}
    ],
    temperature=0.6,
    top_p=0.95,
    max_tokens=1024,
)

message = response.choices[0].message
print("reasoning:", getattr(message, "reasoning_content", None))
print("answer:", message.content)
```

Ornith is a reasoning model. With vLLM's `--reasoning-parser qwen3`, reasoning
tokens are surfaced separately as `reasoning_content`; final answers remain in
`content`.

## Quality and Performance Notes

This upload documents the quantized checkpoint format and provenance. It does
not publish a new retained benchmark table for the quantized model. For
production use, validate quality and end-to-end serving latency against the
BF16 source model on your target workload and runtime.

For original source-model capabilities, benchmark descriptions, and training
context, see the upstream card:
[`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B).

## License and Attribution

The source model is MIT licensed. This derivative quantized artifact keeps the
source license metadata and links to the upstream license file.

If you use the source model, cite the original Ornith release:

```bibtex
@misc{ornith-35b,
    title = {{Ornith-1.0-35B}: Agentic Coding, Open to All},
    url = {https://deep-reinforce.com/ornith_1_0.html},
    author = {{DeepReinforce Team}},
    year = {2026}
}
```