Models & Research

The Linux kernel now contains AI-generated code. Read literally, the new rule would exclude every project that uses it.

It was trained on a single NVIDIA GB200 NVL72. Inference runs on a MacBook Pro M4 Max.

On Moonshot's own Hallu Bench Hard, claim-level hallucinations drop from 23.4 to 8.3 percent. GPT-6 Astra sits at 5.5 percent on the same test.

The model ran for 64 hours largely on its own, from one prompt plus four follow-ups. The project started with a question from an investment research firm.

The hardware line names NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Three pieces a complete serving stack needs are deliberately left out.

Mistral puts realtime transcription at sub-500ms latency across 40+ languages for $0.008 a minute. Speech output runs $13 per million characters.

Staff reckon AI already does 51 percent of the work of building a frontier model. For the end of 2030 their median answer is 95 percent.

The release ships no Jinja-format chat template. In its place come a Python reference implementation and a set of Rust libraries with Python bindings.

A system fixed data errors during training by rolling the run back to an earlier stage. Reflection describes two months with practically zero cluster downtime.

Claude spent 53 hours on a single task and ended up proving new algebraic identities in Lean.

Apple and Meta worked on this for months—two companies that have publicly avoided each other since the 2021 tracking changes. Zuckerberg and Apple chief John Ternus congratulated each other on X.

Google used Argon to pull a six-million-line legacy folder out of Chromium's core page handling, across hundreds of files: months of work in weeks. YouTube's comments section now runs on half the compute.

Codenamed Watermelon, trained at the Prometheus datacenter in Ohio. On MemoryBench, Meta claims it beats Fable, Opus and Astra.

The commitment includes training and onboarding, down to agencies that still have to get a first Genesis project running.

Mathematicians at Cambridge and King's College London found two places where the prose proof and the Lean code diverge.

The model was handed roughly 4,000 problems. Two results fell outside the standard procedure: a zero-free region for the Riemann zeta function, and the Hodge Conjecture for CM abelian varieties.

Haiku 4.5 scored a flat 0.0% on Terminal-Bench 4.0. The 5.5 gets 39.2 — at a tenth of the price.

Behind the generated interfaces sits a compiler that renders the answer while the model is still writing it. Codex and Work are left out.

Strip out the financial profile and let the models read only an email inbox, and much of the gap survives. Gemini read the two finance emails first in 97% of trials.

The reinforcement learning run produced over 100 million rollouts, on 10,500 GB300 GPUs, across four weeks.