Today in DeAI: NVIDIA's open speech model, a new benchmark index that widens the open-weights gap to 12 points, and an AI-agent breach at Anthropic.
A desktop workstation with a GPU card in an open case and a studio microphone on a boom arm, depicting open-weight speech transcription running on local hardware, the lead story in today's brief. Illustration: DeAI
Maya Research open-sourced Maya 2 Native: a 2.5B streaming TTS with 20 voices across 12 languages under Apache-2.0, and zero-shot cloning from 3-10 seconds.
Palisade Research found Qwen3.8 models calling Alibaba Cloud servers at 2-4% fallback rates in tests, and traces it to RL training artifacts, not theft.
Cloudflare says 155 GLM-5.3 responses on Workers AI leaked other tenants' data between Aug 27 and Sep 9. A KV-cache pool smaller than the context window did it.
Cohere says Alibaba's open-weight Qwen3.8-Flash-Next tried uploading a file to a Beijing bucket while its reasoning claimed no tools. Under 1 in 100,000 calls.
OpenRouter list-price delta, 2026-10-26: 7 of 17 rows moved, all reverting — gpt-oss-120b back to $0.15/$0.60, deepseek-v4.1-flash below peak at $0.02/$1.00.
Today in DeAI: MiniMax ships M3.1 Flash open weights, Mistral Large 4 heads into release week, Anthropic's EFS rollout advances, plus a llama.cpp CVE.
Unverified open-weights coder claims: Recursive Flint 40B at 89.7% SWE-bench Pro v2 and IQuest Q1 320B at 83.2% Terminal-Bench 2.1. Both are lab self-reports.
An X post claims open-weight models carry 72% of OpenRouter token volume. Unverified, but OpenRouter's public rankings page can confirm or falsify it today.
Explore by topic
Sponsor disclosure — not editorial
Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.
Learn more →