Model Architecture: Qwen3_5MoeForConditionalGeneration
- Input: Text (+ image/video inputs supported by the base architecture)
- Output: Text
Model Optimizations:
- Activation quantization: FP4
- Weight quantization: FP4
Intended Use Cases: Intended for commercial and research use. Similarly to the base model, this quantized version is intended for agentic coding and general assistant-like chat.
Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
Version: 1.0
Model Developers: RedHat (Neural Magic)

Model Optimizations

This model was obtained by quantizing the weights and activations of deepreinforce-ai/Ornith-1.0-35B to NVFP4 data type. This optimization reduces the number of bits used to represent weights from 16 to 4, significantly reducing GPU memory requirements (by approximately 75%) and increasing inference throughput.

Only weights and activations of the linear operators within the language-model transformer blocks are quantized. The router/gate projections, shared-expert gates, token embeddings, vision tower, and the entire linear-attention (gated deltanet) sub-module are kept at full precision, since they are either not compute-bound Linear layers or are highly sensitive to quantization error. Quantization is performed using the NVFP4 scheme (QuantizationModifier), which targets NVIDIA H100 and Blackwell (sm90+) GPUs, with calibration performed over 256 samples from HuggingFaceH4/ultrachat_200k and moe_calibrate_all_experts=True so that every one of the 256 experts per layer receives calibration signal, not just the experts that are routed to for the sampled tokens. The llm-compressor library is used for quantization.

Hardware requirement: H100 / Blackwell (sm90+) for inference.

Deployment

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

python

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/Ornith-1.0-35B-NVFP4"
number_gpus = 1
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, min_p=0, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [{"role": "user", "content": "Give me a short introduction to large language model."}]
prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)
outputs = llm.generate(prompts, sampling_params)
generated_text = outputs[0].outputs[0].text
print(generated_text)

Creation

Creation details

This model was created with llm-compressor by running the code snippet below. This is the same calibration-based oneshot recipe used upstream for Qwen/Qwen3.5-122B-A10B (the reference example for the Qwen3_5MoeForConditionalGeneration architecture that Ornith-1.0-35B shares), pointed at deepreinforce-ai/Ornith-1.0-35B. Note: unlike Qwen/Qwen3.5-*, the deepreinforce-ai/Ornith-1.0-35B checkpoint does not ship separate MTP (multi-token-prediction) tensors, so the save_mtp_tensors_to_checkpoint step used upstream is omitted here.

python

import torch
from datasets import load_dataset
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.utils import load_context

# NOTE: This example requires transformers >= v5

MODEL_ID = "deepreinforce-ai/Ornith-1.0-35B"

# Load model.
with load_context(Qwen3_5MoeForConditionalGeneration):
    model = Qwen3_5MoeForConditionalGeneration.from_pretrained(MODEL_ID)
processor = AutoProcessor.from_pretrained(MODEL_ID)

recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    ignore=[
        "re:.*lm_head",
        "re:visual.*",
        "re:model.visual.*",
        "re:.*mlp.gate$",
        "re:.*embed_tokens$",
        "re:.*shared_expert_gate$",
        "re:.*linear_attn.*",
    ],
)

NUM_CALIBRATION_SAMPLES = 256
MAX_SEQUENCE_LENGTH = 4096

ds = load_dataset(
    "HuggingFaceH4/ultrachat_200k",
    split=f"train_sft[:{NUM_CALIBRATION_SAMPLES}]",
)
ds = ds.select_columns(["messages"])
ds = ds.shuffle(seed=42)


def preprocess_function(example):
    messages = [
        {"role": m["role"], "content": [{"type": "text", "text": m["content"]}]}
        for m in example["messages"]
    ]
    return processor.apply_chat_template(
        messages,
        return_tensors="pt",
        padding=False,
        truncation=True,
        max_length=MAX_SEQUENCE_LENGTH,
        tokenize=True,
        add_special_tokens=False,
        return_dict=True,
        add_generation_prompt=False,
    )


ds = ds.map(preprocess_function, batched=False, remove_columns=ds.column_names)


def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}


# Apply quantization.
oneshot(
    model=model,
    recipe=recipe,
    dataset=ds,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    moe_calibrate_all_experts=True,
    data_collator=data_collator,
)

# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
model.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)