Qwen3.8-2.4T-A95B Release Notes

Qwen3.8-2.4T-A95B brings Qwen-Max-class capabilities to open release

Qwen3.8-2.4T-A95B is the most capable generation in the Qwen open-model family, delivering substantial gains in coding, professional research, and long-horizon agentic tasks. Built on the architectural foundation of Qwen3.5, this model is designed to handle complex, multi-step tasks with high reliability and stronger autonomous planning.

Model Architecture and Specifications

Qwen3.8-2.4T-A95B is a Causal Language Model utilizing a Mixture of Experts (MoE) architecture to balance total parameter count with inference efficiency.

Core Technical Specs

  • Total Parameters: 2.4 Trillion
  • Activated Parameters: 95 Billion
  • Number of Layers: 92
  • Hidden Dimension: 8192
  • Mixture of Experts: 512 total experts, with 10 routed and 1 shared expert activated per token.
  • Expert Intermediate Dimension: 2048
  • Context Length: 262,144 tokens natively, extensible up to 1,010,000 tokens.
  • Vocabulary Size: 248,320 (padded).

Architectural Layout

The model employs a hidden layout of 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)). It features 128 linear attention heads for V and 16 for QK in the Gated DeltaNet, and 64 attention heads for Q and 4 for KV in the Gated Attention.

Performance Benchmarks

Qwen3.8-Max (the official version based on this model) demonstrates competitive performance against frontier models like Claude Opus 4.8, Fable 5, and GPT-5.6 Sol.

Coding and Agentic Performance

Qwen3.8-Max shows significant improvements over Qwen3.7-Max, particularly in software engineering benchmarks:

  • Terminal Bench 2.1: 86.6 (vs. 74.5 for Qwen3.7-Max)
  • FrontierSWE: 73.5 (vs. 40.7 for Qwen3.7-Max)
  • PaperBench: 93.0 (leading the group with 90.5 for GPT-5.6 Sol)
  • AndroidBench: 75.1 (vs. 56.5 for Qwen3.7-Max)

General Capabilities

  • GPQA Diamond: 92.6, matching Fable 5 and exceeding Opus 4.8 (92.0).
  • IFBench: 82.8, outperforming GPT-5.6 Sol (72.7) and Fable 5 (63.5).
  • MRCR v2 256K (8-needle): 92.9, showing strong long-context retrieval capabilities.

Deployment and API Usage

Serving Frameworks

Qwen3.8 is compatible with several high-throughput inference engines:

  • SGLang
  • vLLM
  • TokenSpeed

API Configuration and Reasoning Control

Qwen3.8-2.4T-A95B is a text-only model that requires "thinking mode" for all interactions. Every response begins with reasoning enclosed in <think> tags.

Users can control the depth of reasoning via the reasoning_effort parameter:

  • xhigh (Default): For complex tasks requiring thorough analysis.
  • medium: Balance between accuracy and speed.
  • low: Optimized for speed and cost.

Recommended Sampling Parameters

To achieve optimal results, the following sampling settings are recommended:

  • temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0.

Community Insights and Technical Considerations

Community discussion highlights several critical trade-offs between the open-weight version and the managed Qwen3.8-Max service.

Hardware Requirements and Quantization

Due to its 2.4T parameter count, the full BF16 model is estimated to be approximately 4.9TB in size. Community members noted that serving this model without quantization is prohibitively expensive for most users.

"The 1bit quant model is at an astonishing 397GB... This literally puts Opus 4.5 performance level into a machine a normal person could buy."

Feature Disparity

Users pointed out that the open-weight version lacks several features present in the managed Qwen3.8-Max version, including:

  • Vision Support: The open-weight model is text-only.
  • Context Window: The open-weight version is capped at 250k tokens, whereas the managed version supports 1M tokens by default.
  • Non-thinking support: The managed version allows for responses without the mandatory reasoning phase.

Licensing

The model is free for internal use or for entities with less than $50M annual revenue. Limitations apply to those exceeding this threshold, particularly for services targeting coding or productivity agents.

Sources

Related

  • Dispatch
    Qwen3.8-Max release notes / what's new

    Qwen has released Qwen3.8-Max, a 2.4 trillion parameter model designed for autonomous coding, professional workflows, and long-horizon tasks, with open weights arriving next week.

  • Dispatch
    Qwen3.8-Max Release: 2.4T‑Parameter Model, Open Weights Next Week, and Broad Autonomous Capabilities

    Qwen3.8‑Max, a 2.4 trillion‑parameter model with 95 B active parameters, is now available via QwenCloud and will have its weights open‑sourced next week, delivering strong gains in coding, real‑world work, long‑horizon tasks, and multimodal agents.

  • Dispatch
    Qwen 3.8 27B Release Notes

    Qwen 3.8 27B is a compact, dense vision-language model featuring flexible thinking control and native image/video understanding, delivering performance competitive with frontier models in coding and agentic tasks.

  • Dispatch
    Qwen3.8-27B Release Notes

    Alibaba Qwen has released Qwen3.8-27B, a native multimodal dense model with 27B parameters that outperforms Qwen3.7-Plus and rivals Claude Opus 4.6 Max in coding and office workflows.

  • Dispatch
    Qwen3.6-35B-A3B Release Notes

    Qwen has released Qwen3.6-35B-A3B, an open-source mixture-of-experts model with 35 billion total parameters and 3 billion active parameters that rivals larger dense models in agentic coding and multimodal reasoning.