Qwen open-sources Qwen3.8-2.4T-A95B, its first Max-class model
Wednesday's wire: Alibaba opened the weights on its most powerful model ever, Honor put a robotic arm in a phone and sold 400,000 of them before lunch, and Europe's booksellers started refusing suspicious bulk orders.
Alibaba's Qwen team open-released Qwen3.8-2.4T-A95B on Wednesday — the first Qwen-Max-class model to ship as open weights, with 2.4 trillion total parameters and 95 billion active in a 512-expert mixture-of-experts architecture. The Hugging Face page went live on the exact schedule the team signaled two weeks ago; a smaller Qwen3.8-27B is promised to follow. Alibaba's own benchmarks put the open model at the top of the frontier: 93.0 on PaperBench, ahead of Claude Opus 4.8 (80.3), Claude Fable 5 (88.8) and GPT-5.6 Sol (90.5), with an 86.6 on Terminal Bench 2.1 that trails only Sol among the models listed. The weights are text-only with thinking mode mandatory, carry a native 262K-token context that extends past a million, and ship with reasoning_effort controls that trade depth for cost.
The significance is less the benchmark table than the precedent. This is the strongest signal yet that China's open-weight labs will keep shipping frontier-scale models anyone can self-host — and it lands as Alibaba reportedly plans revenue-sharing terms for large commercial users, a first for the family and a sign the "open" era is getting commercial. We covered the model's debut two weeks ago, when it became the first open-weight model to top the agentic index — Qwen3.8 Max tops agentic index, OpenAI opens up. What changed today is that anyone with a rack can now run it.
Honor's Robot Phone went on sale in China at ¥9,999 (about $1,400) with pre-orders past 400,000 — a record for the brand and the first real sales signal for a phone with a robotic arm. The device packs a four-axis titanium gimbal that extends out of the body to track faces during video calls, auto-frame shots, and even dance to music; the YOYO agent runs on a foundation model the company says exceeds 300 billion parameters, with a lifetime-free AI membership bundled in. Honor paired the arm with an ARRI-tuned 200-megapixel gimbal camera, and it open-sourced a robot kit so owners can 3D-print shells and bolt the phone onto hobby robots.
Honor built the robot team from scratch just twelve months ago, and CEO Li Jian is pitching the phone as the start of a category — a device that stops taking orders and starts perceiving, deciding and acting. Whether 400,000 pre-orders turn into a durable product line or a flagship novelty, the bet is telling: the next phone wars may be fought over embodied hardware and agentic software, not just chips and screens.
Europe's booksellers are starting to refuse bulk orders they suspect will be destroyed to train AI, after an Irish shop flagged a "bananas" purchase of 5,000 obscure titles. Kennys Bookshop told the Irish Times it will decline orders likely destined for AI training, calling the spree "terrifying our industry" — the first public pushback from the trade since reports that AI firms are buying and pulping books for training data. The practice was exposed by a 2025 lawsuit revealing Anthropic destroyed millions of print books (under a codename, "Project Panama"), and follow-up reporting from the Telegraph and 404 Media documented bulk buying of rare and out-of-print titles.
The counterpoint comes from the Internet Archive, which digitizes even its most fragile books one page at a time, never cutting bindings — slower and pricier, but the existence proof that destruction is a choice, not a requirement. Booksellers refusing orders won't stop AI companies from sourcing text, but it raises the reputational cost of the cheapest path, and pressure from the trade is one of the few forces that has made labs promise to preserve rare books at all.
What to watch: whether Alibaba's revenue-share rate shows up on the Qwen3.8 license, and whether Honor's robot phone becomes a category or a curiosity.
An open 2.4-trillion-parameter model you can self-host — does that change how you think about frontier labs' moats? Tell us in the comments.
