Qwen3.8 Max
Alibaba Qwen3.8 Max: 2.4T sparse MoE, 95B active, 1M context. Excels in autonomous coding, research reproduction, and multimodal understanding of documents.
Model Overview
Capabilities, design details, and architectural traits
Qwen3.8-Max - First Open-Weight Max-Class Model with Long-Horizon Autonomous Coding
Qwen3.8-Max is Alibaba's most capable Qwen model to date and the first Max-class model the company has committed to open-sourcing. Built on the Qwen 3.5 architecture, it combines a Sparse Mixture-of-Experts design with a hybrid attention mechanism, totaling 2.4 trillion parameters while activating only 95 billion per query.
| Trait | Detail |
|---|---|
| Sparse MoE at frontier scale | 2.4T total parameters, 95B active, balancing massive capacity with inference efficiency |
| 1-million-token context | Ingests hundred-page documents, full television series, or 100-hour livestreams into searchable knowledge bases |
| Self-evolving autonomous coding | Ran a 16-day fully autonomous coding project, building the open-sourced oh-my-cli harness with 265 commits and 127 PRs - no human intervention |
| Research reproduction and improvement | Starting from only a paper, reproduced six findings then engineered novel methodologies that outperformed the original work over ~5 days |
| First open-weight Max-class release | Model weights scheduled for public release, a first for Alibaba's Max-tier models |
Feedback-Loop Self-Evolution
A defining behavior of Qwen3.8-Max is its reliance on self-evolving feedback loops rather than fixed plans. In coding, it builds an execution loop that normalizes requirements into issues, dispatches them through a state machine, runs tests, and routes failures back for re-verification. In research, it iterates experiment after experiment. In competitions, it climbs leaderboards submission after submission.
Multimodal Knowledge Construction
The model transforms static visual inputs - documents, video, screenshots - into interactive knowledge structures. It can rebuild software applications from screenshots and convert 2D architectural floor plans into 3D renderings, extending multimodal understanding beyond passive recognition into active construction.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Alibaba | Anthropic | Anthropic |
| Release Date | August 3, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 1M | 1M | 1M |
| Pricing (per 1M tokens) | |||
| Input Pricing | $2 Best Input Pricing | $4 | $2 Best Input Pricing |
| Output Pricing | $6 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimagefilevideo | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 40.2 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 71.8 | - | - |
| Agentic Index | 49.1 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Alibaba
Other models by Alibaba
Top AI Models
Leading alternatives by intelligence score
