Qwen3.8-2.4T-A95B (open weights)
Open weights of a 2.4T-parameter, 95B-active MoE: Terminal Bench 2.1 86.6 and SWE-bench Pro 67.7, under a Qwen3.8-Max license.
92 layers, 512 experts (10 routed plus 1 shared), Gated DeltaNet plus gated attention, thinking mode always on with xhigh/medium/low effort and preserved reasoning. Brings a Max-class model to open release, but not under Apache 2.0.
- Date
- Wednesday, 12 August 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Terminal Bench 2.1 | 86.6 SWE-bench Pro 67.7; context 262,144 native, ~1.01M extended | company |
GitHub README dates the release 2026-08-12 (HF repo created 2026-08-08). The Hugging Face card says 'qwen3.8-max license'; Wikipedia reports it requires a commercial agreement above $50M revenue, which I did not verify in the license text. Caixin called it the first flagship Qwen model released publicly.
Sources
This record was checked against its sources on 6 October 2026. How we check