Qwen3.8-2.4T-A95B

Model Ecosystem

Model Tree

Finetunes

• 4 models fine-tuned from Qwen/Qwen3.8-2.4T-A95B
• View finetunes

Quantizations

• 29 quantized models available
• Supported tools: llama.cpp, LM Studio, Jan, Ollama
• View quantizations

Spaces

• 6 Spaces using Qwen/Qwen3.8-2.4T-A95B

Collection

Part of the Qwen3.8 collection:

• 4 items • Updated 11 days ago • 414 followers
• View collection

Community

• 38 discussions on HuggingFace
• 99.9k followers
• 18,893 downloads last month

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}

Source

• Blog post: Qwen3.8-Max
• HuggingFace: Qwen/Qwen3.8-2.4T-A95B

Thinking Control

Overview

Qwen3.8-2.4T-A95B is a text-only model that requires thinking mode for all interactions. Thinking cannot be disabled. Every response automatically begins with reasoning enclosed in <think>\n...\n</think> before the final output.

reasoning_effort Levels

Reasoning depth can be tuned to balance accuracy and cost:

preserve_thinking

• Enabled by default for all workloads.
• Retains reasoning context from historical messages, providing the best out-of-the-box experience.
• Can be passed via extra_body with chat_template_kwargs in self-hosted inference, or directly in Qwen Cloud API.

API Parameters

extra_body={
    "chat_template_kwargs": {
        "enable_thinking": True,   # on by default; should not be turned off
        "preserve_thinking": True  # on by default
    }
}
reasoning_effort="xhigh"  # xhigh by default; supported: xhigh, medium, low

Best Practices

Sampling Parameters

Recommended sampling parameters for optimal generation:

Notes on presence_penalty

• Can be adjusted between 0 and 2 to reduce endless repetition.
• Higher values may occasionally result in language mixing and slight performance decrease.
• Support for sampling parameters varies by inference framework.

Output Length Configuration

For agentic tasks, allocate sufficient output length within the 1M context:

These settings provide necessary capacity for complex reasoning while ensuring ample space for high-quality final deliverables.

API Usage

Key Characteristics

• Text-only model: Multimodal inputs are not supported.
• Thinking mode required: Thinking cannot be disabled. Every response automatically begins with reasoning enclosed in <think>\n...\n</think> before the final output.
• preserve_thinking: Enabled by default for all workloads.

Chat Completions API (OpenAI-compatible)

from openai import OpenAI
## Configured by environment variables
client = OpenAI()

messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}]

completion = client.chat.completions.create(
    model="Qwen/Qwen3.8-2.4T-A95B",
    messages=messages,
    extra_body={
        "chat_template_kwargs": {
            "enable_thinking": True,  # on by default; should not be turned off
            "preserve_thinking": True, # on by default
        },
    },
    reasoning_effort="xhigh",  # xhigh by default; supported: xhigh, medium, low
    stream=True,
    stream_options={"include_usage": True},
)

reasoning_content = ""
answer_content = ""
is_answering = False

for chunk in completion:
    if not chunk.choices:
        print(chunk.usage)
        continue
    delta = chunk.choices[0].delta
    if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None:
        reasoning_content += delta.reasoning_content
    if hasattr(delta, "content") and delta.content:
        is_answering = True
        answer_content += delta.content

Qwen Cloud API

When using Qwen Cloud APIs, pass extra_body directly (not nested under chat_template_kwargs):

extra_body={"enable_thinking": True, "preserve_thinking": True}

Deployment Options

Supported Inference Frameworks

Qwen3.8-2.4T-A95B is compatible with multiple popular inference frameworks:

SGLang

• Website: https://www.sglang.io/
• Cookbook: Qwen3.8 Cookbook

vLLM

• Website: https://vllm.ai/
• Recipe: Qwen3.8 Recipe

TokenSpeed

• Website: https://lightseek.org/tokenspeed/
• Recipe: Qwen3.8 Recipe

Managed Cloud Service

For users seeking managed, scalable inference without infrastructure maintenance, the official Qwen API service is provided by Qwen Cloud.

Recommendations

• For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, vLLM, or TokenSpeed are recommended.
• Inference efficiency and throughput vary significantly across frameworks — use the latest framework versions for optimal performance and compatibility.

Benchmark Results

Coding Agent Benchmarks

General Agent Benchmarks

General Capabilities Benchmarks

HuggingFace Evaluation Results

• DeepSWE: 56.6
• GPQA Diamond: 92.6
• SWE-bench Pro: 67.7
• WildClawBench Overall: 56.2
• Terminal Bench 2.1: 86.6
• HLE: 43.6

Context Length

Native Context Length

• 262,144 tokens natively supported.

Extended Context Length

• Extensible up to 1,010,000 tokens (approximately 1M).

Notes

• Qwen3.8-Max (the official hosted version) offers 1M context length by default.
• For long-context tasks, the model supports RoPE-based position scaling.
• The extended context is particularly useful for agentic workflows requiring extensive reasoning and output space.

Model Architecture

Overview

• Type: Causal Language Model
• Training Stage: Pre-training & Post-training
• Total Parameters: 2.4T (2,400,000,000,000)
• Activated Parameters: 95B (95,000,000,000)
• Hidden Dimension: 8192
• Token Embedding: 248,320 (Padded)
• Number of Layers: 92
• Tensor Type: BF16

Hidden Layout

The model uses a repeating pattern of 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)), alternating between linear attention and full attention layers.

Gated DeltaNet (Linear Attention)

• Number of Linear Attention Heads: 128 for V and 16 for QK
• Head Dimension: 128

Gated Attention (Full Attention)

• Number of Attention Heads: 64 for Q and 4 for KV
• Head Dimension: 256
• Rotary Position Embedding Dimension: 64

Mixture of Experts (MoE)

• Number of Experts: 512
• Number of Activated Experts: 10 Routed + 1 Shared
• Expert Intermediate Dimension: 2048

Multi-Token Prediction (MTP)

• Trained with multiple steps

LM Output

• 248,320 (Padded)

Qwen3.8 Highlights

Qwen3.8 is the most capable generation in the Qwen open-model family to date, bringing a Qwen-Max-class model to open release for the first time. Built on the architectural foundation of Qwen3.5, it delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.

Core Enhancements

• Core Capabilities: Comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks.
• Agent Execution: Stronger autonomous planning and better handling of environment feedback, leading to more reliable end-to-end task completion.
• Downstream Compatibility: Broader support for popular harnesses and development tools, making it easier to integrate into existing stacks.
• Flexible Thinking Control: Reasoning depth can be tuned with reasoning_effort, and reasoning context from historical messages is retained via preserve_thinking.

Qwen3.8-Max

Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with additional features:

• Vision input support
• Non-thinking support
• 1M context length by default
• Official built-in tools

Blog post: Qwen3.8-Max