Qwen3.8-Max Released: 2.4T Parameters, Open Weights Next Week

What was released

On 2026-08-03, Alibaba's Qwen team published a blog post introducing Qwen3.8-Max, the largest model in the Qwen family to date. The publisher describes the model as a sparse Mixture-of-Experts architecture with hybrid attention, scaling to 2.4 trillion total parameters with 95 billion parameters activated per token, and a native 1-million-token context window.

The publisher also announced that the open-weight release of Qwen3.8-Max and the smaller Qwen3.8-27B is scheduled for the following week, distributed via HuggingFace and ModelScope. Until the weights ship, developers can access the model through the QwenCloud API platform.

What the publisher documents

The following details are taken from the cited publisher blog and API documentation. They describe what the publisher states about the model, not independent findings of this site.

• Architecture: Sparse Mixture-of-Experts with hybrid attention mechanisms; 2.4T total parameters, 95B active per token (per the publisher blog).
• Context window: 1,000,000 tokens native, including visual understanding (per the publisher announcement).
• Benchmarks (publisher-reported):PaperBench 93.0, WideSearch 81.9, Agent's Last Exam 52.4, IFBench 82.8, GPQA Diamond 92.6, BabyVision 82.0, OSWorld-Verified 86.1. CodeArena global rank 4.
• API access: QwenCloud platform with both OpenAI-compatible and Anthropic-compatible endpoints. Integrates with Claude Code, Codex, Qoder CLI, and OpenClaw.
• Reasoning control: Publisher documents a reasoning_effort parameter with three levels (xhigh, medium, low) for adjusting inference depth.
• Pricing (China): ¥12 / M input tokens, ¥36 / M output tokens, ¥1.5 / M cache-hit tokens (per publisher pricing page).
• Open weights: Scheduled for release the week of 2026-08-03 on HuggingFace and ModelScope. A 27B variant (Qwen3.8-27B) will be open-sourced alongside.

Long-horizon autonomous coding case

The publisher blog highlights three autonomous coding demonstrations. The most cited is a 10+ day run in which Qwen3.8-Max built a self-evolving agent harness (“oh-my-cli”) from an empty folder, accumulating 265 commits, 127 pull requests, and 151 issues by 2026-07-30. The full trace is published in the publisher's GitHub repository (qwen-code-dev-bot/oh-my-cli).

The publisher also reports reproducing and then improving upon the paper “Unified Data Selection for LLM Reasoning”, achieving a 2.7-point gain on the AIME24 benchmark versus the paper's original method, and reaching 85.3% accuracy in the WWW2025 multimodal intent recognition challenge, beating 87% of human teams.

These results are publisher-reported. This site has not independently reproduced them.

What developers can do next

1. Read the official blog on qwen.ai for the full benchmark tables and architecture notes.
2. Try the API on QwenCloud to evaluate the model before the open-weight release. Both OpenAI-compatible and Anthropic-compatible endpoints are documented.
3. Watch HuggingFace and ModelScope for the open-weight release during the week of 2026-08-03. Confirm the exact license terms in the model card before commercial use.
4. Compare with peer open-weight models such as LongCat-2.0 (1.6T parameters, MIT license, 1M context), DeepSeek V4-Flash, and Kimi K3 (2.8T parameters, open-sourced 2026-07-27). Independent comparisons from this site are linked in the related pages below.

Source notes and limitations

• This site has not independently tested Qwen3.8-Max. All capability and benchmark descriptions are sourced from the cited publisher blog and API documentation.
• The open-weight release is scheduled but not yet available as of publication. Confirm availability on HuggingFace and ModelScope before downloading.
• Pricing and API behaviour may change. Verify the current details on the publisher's pricing page before deployment.
• This page is published by an independent third-party site and is not affiliated with or endorsed by Alibaba, Qwen, or Anthropic.

Related pages

• LongCat-2.0 model profile

  Compare with Meituan's open-source flagship: 1.6T parameters, 1M context, MIT license.

• Benchmark evidence index

  How this site labels publisher-reported vs third-party metrics for open-source LLMs.

• LongCat-2.0 vs GPT-5.5 comparison

  A source-based head-to-head covering architecture, pricing, and licensing.

Sources

• Qwen3.8-Max: A New Bar for Coding and Cowork — Official Blog

  Publisher documentation · Published 2026-08-03 · Accessed 2026-08-03

  Official Qwen team announcement. Contains architecture summary (2.4T total parameters, 95B active, MoE with hybrid attention), benchmark scores, autonomous coding case studies, and API availability details.

• QwenCloud API Platform

  Publisher documentation · Accessed 2026-08-03

  Publisher API platform. Confirms OpenAI-compatible and Anthropic-compatible endpoints, reasoning_effort parameter (xhigh/medium/low), and integration with Claude Code, Codex, Qoder CLI, and OpenClaw.

• Qwen3.8-Max on HuggingFace (upcoming)

  Primary source · Accessed 2026-08-03

  Publisher HuggingFace organization. Open weights for Qwen3.8-Max and Qwen3.8-27B scheduled for release during the week of August 3, 2026, per publisher announcement.

• Qwen3.8-Max on ModelScope (upcoming)

  Publisher documentation · Accessed 2026-08-03

  Alibaba ModelScope model repository. Confirmed as a secondary mirror for the open-weight release alongside HuggingFace.