Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-05 · high
Community reports say usage is restricted in the US, EU, UK and Korea, breaking Qwen's usual Apache 2.0 license, though Alibaba hasn't confirmed terms yet.
Why now: Alibaba announced the model on August 3, and as of this coverage on August 5, no license text has surfaced for the weights due next week.
Alibaba released Qwen 3.8 Max on August 3, a 2.4 trillion-parameter model scoring 87.3% on SWE-bench, and shipped it without license terms. Terminal-Bench 2.1 scores jumped from 61.0 for the previous Qwen 3.7 Max to 67.4, a real gain for open-weight coding models.
It ranks #4 on the Frontend Code Arena, at 1,668 Elo. Among open-weight models on the Vals Index, it's #2, at 66.1. Access through Alibaba's API runs $2 per million input tokens and $6 per million output tokens, with cached tokens at $0.25 per million. Context runs to 1 million tokens, with a 128k max output.
The weights for Qwen 3.8 Max and a smaller companion, Qwen3.8-27B, are due on Hugging Face and ModelScope "next week," per Latent Space's AINews. Alibaba named no license at announcement. Community reports cited by AINews say usage is restricted in the US, EU, UK and Korea. That would break from the Apache 2.0 terms that covered prior Qwen releases.
Apache 2.0 is what made Qwen a real option for US teams fine-tuning open models instead of paying OpenAI or Anthropic per token. If next week's license excludes the US, EU, UK and Korea, Qwen 3.8 Max stops being that option for those developers. A benchmark chart isn't a permission slip. Alibaba announced this August 3, and two days on, no license text has shown up. I wouldn't plan around these weights until it does.
Each link below shares sources, entities, or timing with this story.
Moonshot AI dropped Kimi K2.6 today and the numbers are hard to ignore. One trillion parameters total, 32 billion active per token across 384 experts, 256K context window, and native multimodal input. It scores 58.6 on SWE-Bench Pro versus GPT-5.4's 57.7 and Claude Opus 4.6's...
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
Alibaba's Tongyi Lab released it on August 14: 27.78B dense parameters, text/image/video input, native 262,144-token context extensible to 1M through YaRN (AI Release Tracker). Reported scores include LiveCodeBench 90.3%, GPQA Diamond 89.2%, and Terminal-Bench 2.1 at 73.0, up...
Poolside AI released two models that change the math on local coding agents. Laguna M.1 is a 225B total / 23B active MoE model scoring 72.5% on SWE-bench Verified. Laguna XS.2 is a 33B total / 3B active model scoring 68.2% on the same benchmark, 44.5% on SWE-bench Pro, and 30....
The Hugging Face page is marked "Upcoming release" with no model card, license, architecture details, context length or benchmarks, after Alibaba promised both Qwen3.8-Max and the 27B weights for the week of August 10. A ModelScope countdown pointed at August 15. Unsloth signa...
Alibaba released Qwen3.6-27B on April 22. Dense architecture. Open weights. 77.2% on SWE-bench Verified, within 3.7 points of Claude Opus 4.6. On SkillsBench, it scores 48.2% versus its own 397B MoE predecessor's 30.0%. That's a 77% improvement with 14.8x fewer parameters. Let...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.