Instructions to use RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8")
model = AutoModelForCausalLM.from_pretrained("RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8

SGLang

How to use RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 with Docker Model Runner:
```
docker model run hf.co/RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8
```

Qwen3.8-2.4T-A95B-NVFP4-FP8

Model Overview

Model Architecture: Qwen3_5MoeForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: Mixed (FP4 MoE experts, FP8 attention)
- Activation quantization: Mixed (FP4 / FP8)
Release Date: 2026-08-14
Version: 1.0
Model Developers: RedHatAI

This model is a quantized version of Qwen/Qwen3.8-2.4T-A95B. It was evaluated on several tasks to assess its quality in comparison to the unquantized model.

Model Optimizations

This model was obtained by applying mixed-precision quantization to Qwen/Qwen3.8-2.4T-A95B: the MoE expert linear layers use FP4 (NVFP4) weights and activations, while the attention linear layers use FP8 (W8A8 block) weights and activations, ready for inference with vLLM.

This optimization reduces the number of bits per parameter in the quantized layers from 16 to 4 or 8 depending on the layer, reducing the disk size and GPU memory requirements by approximately 50-75%.

Only the weights and activations of the linear operators in the MoE experts and attention blocks are quantized using LLM Compressor.

Deployment

vLLM Serving

vllm serve RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8 \
    --data-parallel-size 8 \
    --enable-expert-parallel 8 \
    --reasoning-parser qwen3 \
    --max-num-seqs 140

NOTE: Because of the need for data parallelism, this model may use more model memory for replicated attention layers. This can lead to less memory for kv cache and cache preemption when handling large concurrency. For high concurrency tasks, consider using RedHatAI/Qwen3.8-2.4T-A95B-NVFP4.

Creation

This model was created by applying LLM Compressor with the NVFP4 (MoE) + FP8 block (attention) mixed-precision scheme, exported in compressed-tensors format.

# NOTE: to use a custom dataset, see examples/custom_dataset_example.py
from compressed_tensors.offload import init_dist
from compressed_tensors.quantization.quant_scheme import (
    FP8_BLOCK,
    NVFP4,
    QuantizationScheme,
)
from transformers import AutoModelForCausalLM, AutoTokenizer

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

# Select model and load it.
init_dist()
model_id = "inference-optimization/Qwen3.8-1.0B-A0.6B"  # Qwen/Qwen3.8-2.4T-A95B
with load_context():
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map="auto_offload",
        max_memory={},
        offload_folder="offload_folder",
    )
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Create recipe
recipe = [
    QuantizationModifier(
        config_groups={
            "attention": QuantizationScheme(
                targets=[
                    r"re:.*self_attn\..*",
                    r"re:.*linear_attn.(in_proj_qkv|in_proj_z|in_proj_b|in_proj_a|out_proj)$",
                ],
                **FP8_BLOCK,
            ),
            "mlp": QuantizationScheme(
                targets=[r"re:.*mlp\..*"],
                **NVFP4,
            ),
        },
        ignore=[
            "re:.*lm_head",
            "re:.*mlp.gate$",
            "re:.*shared_expert_gate.*",
        ],
    ),
]

# Apply algorithms.
oneshot(
    model=model,
    dataset="perfectblend",
    splits="train[:512]",
    recipe=recipe,
    max_seq_length=2048,
    num_calibration_samples=1024,
    pipeline="sequential",
)

# Save to disk compressed.
SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-NVFP4-FP8"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

Evaluation

This model was evaluated on GPQA Diamond, served with vLLM (OpenAI-compatible API). Recovery is computed against the unquantized model.

Accuracy

Category	Benchmark	Qwen/Qwen3.8-2.4T-A95B	RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8	Recovery
Reasoning	GPQA Diamond	92.6	93.1	100.54%

Downloads last month: 653

Safetensors

Model size

1.4T params

Tensor type

BF16

F8_E4M3

Model tree for RedHatAI/Qwen3.8-2.4T-A95B-NVFP4-FP8

Base model

Qwen/Qwen3.8-2.4T-A95B

Quantized

(32)

this model