Alibaba's Qwen 3.8 Max (2.4T) goes open-weight, ranks #2–4 on major coding benchmarks

Alibaba Qwen announced Qwen3.8-Max on August 3, 2026, positioning it as the company's "most capable model to date" and a flagship focused on autonomous coding, long-horizon agent work, and native multimodal reasoning. The model is a 2.4T-parameter sparse architecture with approximately 95B active parameters per token, yielding a ~4% mixture-of-experts activation ratio. It supports a 1M-token context window and up to 128k tokens of output.

The API pricing is $2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million cached tokens—a reduction from Qwen3.7-Max pricing ($2.50 / $7.50). Alibaba committed to releasing open weights for both Qwen3.8-Max and Qwen3.8-27B "next week," making the flagship model's weights freely available rather than API-only.

Alibaba's own claims for the model are notably aggressive: 10+ days of autonomous coding with a public GitHub trace, 500+ turns for chip design optimization, 365 days of e-commerce strategy execution, and native multimodal planning where vision is integrated into the execution loop rather than serving merely as an input channel. The company simultaneously pushed availability across multiple surfaces—Qwen Studio, API, Command Code, and Venice—with rapid partner support from Baseten, Hermes Agent, and Command Code.

Third-party benchmarks confirmed strong placements. Arena reported Qwen3.8-Max at #4 in Frontend Code Arena with 1,668 Elo (trailing Claude Opus 5 Max at 1,705 and Kimi K3 Max at 1,676), and #2 in Vision Arena at 1,305 Elo, 13 points behind Claude Fable 5 High. Vals AI ranked it #2 among open-weight models and #10 overall out of 43 models with a score of 66.1, matching Claude Opus 4.7 (also 66.1) while costing roughly 2.3× less per test ($2.68 vs. $6.17). On SWE-bench, Qwen3.8-Max achieved 87.3%, ahead of GPT-5.5 (82.6%) and GLM-5.2 (83.3%), though behind Claude Opus 4.8 (89.2%). Terminal-Bench 2.1 showed 67.4, up from 61.0 for Qwen3.7-Max, representing an 8.6-point gain over ~2.5 months.

However, important caveats emerged in the reaction. Jamin Ball emphasized that pricing comparisons are misleading because vanilla token prices ignore token efficiency and because these models are operationally prohibitive for local deployment: the 2.4T Qwen requires at least 8 H100/B200 GPUs and over 1TB of memory just to load weights, effectively identical to Kimi K3's infrastructure demands. Users flagged a more serious concern: potential geographic restrictions in the license terms that appear to cover the US, EU, UK, and Korea, with one observer (OstrisAI) reading the terms as forbidding even download from the US. Alibaba did not issue a clarifying tweet on licensing within the dataset, leaving the "open weights" claim ambiguous—the model may be restricted from commercial deployment, export, or even download in major regions.

The release was widely interpreted as a strategic pivot by Alibaba away from API-only exclusivity toward ecosystem influence, driven by competition from DeepSeek, Kimi, and other Chinese open models that had already weakened the premium of keeping top-tier systems closed. The real adoption wave, observers suggested, may come not from the 2.4T flagship but from the smaller Qwen3.8-27B, which practitioners expect to inherit post-training or distilled capabilities while remaining practical for developers on consumer hardware.