Alibaba plans revenue-share on next open-weight Qwen for big users
Free for tinkerers; scale a business on the weights and Alibaba reportedly wants a cut. Plus: DeepSeek V4 Flash tops HN, MiniMax H3 video, Vercel Skill Packs.
Copy markdown
The revenue-share catch
Alibaba reportedly plans to ask large commercial users of its next open-source Qwen model to share revenue — a break from the no-strings Apache-2.0 terms Qwen built its following on. Personal and small-scale use would stay free, but if you're shipping a product on Qwen weights at scale, budget for a license conversation.
The weights land within days
The new terms reportedly attach to the next open release: Qwen3.8-Max (2.4T MoE, 1M context) plus a 27B, both due on Hugging Face and ModelScope this week. The license may arrive stapled to the download — read it before you wire Qwen into anything production.
The open model builders are actually running: DeepSeek V4 Flash
MIT-licensed, 284B total / 13B active, 1M context — and a post-training pass alone pushed the 0731 build past DeepSeek's own 1.6T V4-Pro on the Artificial Analysis Intelligence Index (50 vs 44), with Terminal-Bench 2.1 near 79%. At $0.14/$0.28 per Mtok, it's the cheap, self-hostable agent brain topping Hacker News today.
Open-weight video with a geo-lock: MiniMax H3
MiniMax's Hailuo 3.0 tops a video-editing arena and ships open weights (~42 GB) that generate and edit in one model — but the license bars local deployment in the US, EU, UK, and Korea. Translation: fine to build on via API, off-limits to self-host if that's where you are.
Vercel adds Skill Packs and Seedance 2.5
You can now bundle and batch-update agent skills from the skills.sh directory, and ByteDance's Seedance 2.5 (up to 30-second clips) is live on Vercel AI Gateway with no token markup. A weekend unlock: add text-to-video to an app without leaving your deploy pipeline.