Qwen 3.8-2.4t-a95b

Terminal-Bench 2.1: 86.6% (vendor) / 81.3% (AA independent), SWE-bench Pro: 67.7%, DeepSWE 1.1: 56.6%, PaperBench: 93.0% (#1), GPQA Diamond: 92.6%, FrontierSWE: 73.5%, OSWorld-Verified: 86.1%, OmniDocBench 1.5: 92.1%, IFBench: 82.8%, HLE: 43.6%, JobBench: 53.4% | Alibaba's most capable model and the first Qwen-Max released as open weights (GA Aug 3, 2026; open weights Aug 12, 2026). 2.4T-parameter multimodal reasoning system — the largest in the Qwen family. 1M-token context window (256K native, extensible to ~1.01M). Natively multimodal: text, image, and video input. Configurable reasoning effort (low/medium/xhigh). **Leads PaperBench at 93.0% (#1, ahead of GPT-5.6 Sol at 90.5). Beats Opus 4.8 on Terminal-Bench 2.1 (86.6% vs 84.6% vendor-reported). Strong computer-use (OSWorld 86.1%) and document understanding (OmniDocBench 92.1%).** Alibaba claims "second only to Fable 5." $2/$6 per MTok — under a quarter of Opus 4.8's output price. Open weights on Hugging Face as `Qwen3.8-2.4T-A95B`. **Caveats: AA independently measured Terminal-Bench at 81.3% (5.3 pts below vendor's 86.6%); SWE-bench Pro at 67.7% trails Fable 5's 80.0%; DeepSWE 1.1 at 56.6% trails all closed frontier peers; open weights require ~1.2 TB VRAM at FP4 — API-only for most users. Vendor-reported benchmarks, limited independent verification.**