Alibaba has launched Qwen3.8-Max, a 2.4-trillion-parameter sparse mixture-of-experts model with 95 billion active parameters per token, and announced that model weights will be open-sourced next week1,2. A smaller Qwen3.8-27B model will also go open-weight.
The release marks a return to form for the Qwen team. After leadership changes and a period of closed API-only releases following what Latent Space described as the "Qwen Exodus," there had been doubt about whether the lab would continue shipping competitive open models.
Qwen3.8-Max is available now via API on Qwen Studio, Command Code, and other surfaces at $2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million cached tokens. The model features a 1-million-token context window and 128,000 max output tokens. It exposes low, medium, and xhigh reasoning-effort modes through its API and is compatible with OpenAI and Anthropic protocols.
Alibaba is positioning the model around long-horizon autonomous work, coding, and multimodal reasoning. Among the headline capability claims: Qwen3.8-Max reportedly ran a self-evolving coding harness over a multi-week autonomous run spanning 10-plus days, with a public GitHub trace. In an autonomous research demonstration, Alibaba said the model rebuilt a complete paper-replication pipeline and ran an iterative research loop over 125 hours, producing a new data-selection method that beat the original paper's benchmark by 2.71 points. In a competitive data-science setting, Alibaba said the model placed in the top 13% against 526 human teams in the WWW2025 Multimodal Dialogue Intent Recognition Challenge within 24 hours.
On hardware design, Alibaba said Qwen3.8-Max executed a complete silicon design flow for a GCD/RSA cryptographic accelerator — from RTL editing through simulation, synthesis, and physical layout — reducing gate count from 8,298 to 678 gates, achieving an 81% die-area reduction, and meeting timing closure at 500 MHz. Alibaba said the model can perform 500-plus turns of chip-design optimization.
Alibaba described the model's vision capability as "native multimodal intelligence" where vision is part of the execution loop rather than just an input channel.
Third-party benchmark results are emerging. Vals reported Qwen3.8-Max scored 87.3% on SWE-bench, 66.1 on the Vals Index, and 67.4 on Terminal-Bench 2.1. The Vals Index score matched Claude Opus 4.7's 66.1, while the per-test cost was $2.68 versus $6.17 for Claude Opus 4.7. Vals also noted an 8.6-point gain on its index compared to Qwen 3.7 Max's score of 57.5, achieved in roughly 2.5 months. Qwen3.8-Max ranked second among open-weight models on the Vals Index. Arena reported the model debuted at number four overall in Frontend Code Arena with 1,668 Elo and number two in Vision Arena with 1,305 Elo. ZhihuFrontier reported benchmark claims of PaperBench 93.0, CoWorkBench 74.8, and WideSearch 81.9.
Baseten, Hermes Agent, and Command Code have confirmed support plans or integrations for Qwen3.8-Max.
ANALYSIS The 2.4T total parameter count with only 95B active per token implies roughly a 4% activation ratio, a design that trades total model size for inference efficiency. Matching Claude Opus 4.7 on the Vals Index at less than half the per-test cost sharpens the competitive pressure on frontier API pricing. The commitment to open-weight both the flagship and the 27B variant next week, if fulfilled, would give the open-source ecosystem access to one of the largest sparse MoE models released to date.