Alibaba Qwen3.8 2.4T A95B: 2.4T open-weight MoE with 95B active params, hybrid linear/full attention, configurable reasoning, 1M context for agentic
Capabilities, design details, and architectural traits
Alibaba's Qwen3.8-2.4T-A95B is the first Qwen-Max-class model released as open weights. It is a sparse mixture-of-experts model, built on the Qwen3.5 architectural foundation and designed for coding, research, professional work, and long-horizon agentic tasks.
The defining architectural feature is a hybrid attention layout that alternates between Gated DeltaNet (linear attention) and Gated Attention (full attention). The 92-layer stack follows a repeating pattern of 23 blocks, each containing 3 Gated DeltaNet layers followed by 1 Gated Attention layer, all paired with MoE. Linear attention replaces the growing KV cache with a bounded recurrent state, keeping compute and memory bounded as context scales toward one million tokens.
| Trait | Detail |
|---|---|
| Hybrid attention layout | 23 × (3 × Gated DeltaNet → MoE → 1 × Gated Attention → MoE), 92 layers total |
| Fine-grained MoE | 512 experts, 10 routed + 1 shared activated per token |
| Configurable reasoning | reasoning_effort (low / high / xhigh) controls depth; preserve_thinking retains reasoning context across messages |
| Multi-Token Prediction | Trained with multiple MTP steps |
| Context window | 262,144 tokens natively, extensible to 1,010,000 |
| First open Qwen-Max-class release | Qwen3.8-Max is the official managed version adding vision input, non-thinking mode, 1M default context, and built-in tools |
The model is designed to carry complex, multi-step tasks through to completion with stronger autonomous planning and better handling of environment feedback. Reasoning depth can be tuned per request - dialed up for multi-step reasoning or dialed down for high-throughput document processing. Day-0 support shipped for vLLM, SGLang, and TokenSpeed, with FP8 and BF16 checkpoints available at launch.
Independent evaluations · Artificial Analysis
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Alibaba | Anthropic | Anthropic |
| Release Date | August 12, 2026 | July 24, 2026 | June 9, 2026 |
| Knowledge Cutoff | - | May 2026 | - |
| Context & Limits | |||
| Context Window | 262K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $2 Best Input Pricing | $5 | $10 |
| Output Pricing | $6 Best Output Pricing | $25 | $50 |
| Modalities | |||
| Inputs | text | textimage | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 57.7 | 63.1 Best Intelligence Index | 62.1 |
| Coding Index | 71.9 | 78.0 Best Coding Index | 76.5 |
| Agentic Index | 57.1 | 59.2 Best Agentic Index | 56.6 |
Graduate-level reasoning and expert Q&A evaluation.
Extremely difficult logical reasoning and knowledge.
Logical reasoning over long context windows.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
