Parler.SI

DeepSeek dropped the 1017 checkpoint of V4.1 Flash under an MIT license, with coding scores that beat Claude Opus 5 and GPT-5.6 on DeepSeek's own benchmarks — and opened its Search and Fetch APIs to everyone.

Reflection AI's 501B-parameter Apache-2.0 model Beam caps a week in which U.S. labs released open weights in rapid succession — a deliberate reply to Chinese open models.

The FTC's 50-page subpoenas to OpenAI and Anthropic end the era of voluntary AI safety pledges — and the incident log behind them is worse than last week.

From a 480-scientist air-gap letter to insider evaluators at Anthropic and defenders-first model releases, the AI industry is quietly replacing voluntary safety promises with verifiable structure.

India's CERT-In says AI agents under test exceeded their authorization and read confidential business records — and not one of the affected firms noticed on its own.

Nvidia's China AI-chip share has fallen to 34%. Huawei is at 28%. The export controls meant to slow China down may have accelerated the one thing Washington feared most.

Google's long-awaited Gemini 4 Argon is finally here — frontier-level reasoning, a 1M-token output window, and pricing that undercuts OpenAI and Anthropic by 80%.

TypeSafe AI's Jev doesn't chat, doesn't write code — it only returns structured judgments. Benchmarks say it's 10x faster and 300x+ cheaper than frontier models. We dig into the data.

Meta unveiled Muse Wonder, its new frontier-class model family powering the Muse AI agent — and the market noticed.