Qwen3.8-Max preview: Alibaba claims "second only to Fable 5" — with no benchmark table
Alibaba's Qwen team previewed Qwen3.8-Max, a 2.4-trillion-parameter multimodal flagship — its first trillion-plus model to handle images, video and documents alongside text — announcing it during WAIC in Shanghai, two days after Moonshot's Kimi K3 launch. The team says the model ranks "second only to Fable 5" among the systems it benchmarked, but the announcement ships with no benchmark table, no model card and no license; no independent evaluation exists yet. The preview is usable today through Qwen's Token Plan subscription at promotional pricing, and open weights are promised without a date. The announcement itself is X-only — nothing on the Qwen blog or feeds; only the pricing page names the model.
Sources & depth
Kimi K3 demand forces Moonshot to pause new subscriptions
Within roughly 48 hours of Kimi K3's July 16 launch, Moonshot AI suspended new consumer subscriptions, saying surging demand had pushed its GPU fleet to capacity. Existing subscribers keep full service; new slots will reopen in batches as the company adds compute, with no timeline given. The full open-weights release remains scheduled for July 27. The rush is the clearest signal yet that K3's claimed frontier parity converted into real load — serving capacity, not model quality, is now the bottleneck for China's top labs.
Sources & depth
OpenAI quietly cuts Codex's GPT-5.6 context window from 372k to 272k
A merged pull request in OpenAI's open-source Codex CLI repo ("Backport refreshed bundled model metadata to 0.144", merged July 18) drops the bundled context window for GPT-5.6 Sol from 372,000 to 272,000 tokens in models.json — and the current catalog now lists 272k across every bundled model, from GPT-5.6 Sol/Terra/Luna down to GPT-5.2. No reason is given in the PR, which frames the change as a catalog refresh backported to the stable 0.144 branch. For agent workloads sized against the old 372k ceiling, that is a silent ~27% cut in usable context on the flagship tier.
Sources & depth
Claude Code now ships on Bun's Rust port
Bun creator Jarred Sumner claimed that Claude Code v2.1.181 (released June 17) and later run on the Rust port of Bun, with about 10% faster startup on Linux — and that barely anyone noticed. Simon Willison verified it on his own installation: running strings against the Claude binary surfaced "Bun v1.4.0", a version that predates any official Bun release, plus 563 embedded Rust source filenames, and a TypeScript preload trick confirmed the runtime version from inside the process. A full runtime rewrite shipping invisibly under one of the most-used developer tools is the quiet kind of infrastructure milestone.
Sources & depth
AI advice made people less accurate — and far more confident
A new preprint from researchers at Milano-Bicocca, ENS and Sapienza had subjects answer visual trivia questions with and without access to an AI model deliberately chosen because it fails those questions. With the AI available, accuracy fell from 27% to 9% while confidence rose from 30% to 76%, and willingness to admit not knowing collapsed from 44% to 3%; monetary incentives barely helped. It is a clean demonstration that a confidently wrong assistant doesn't just fail to help — it actively suppresses the checking people would otherwise do.
Sources & depth