vLLM says open‑source TPU kernel runs Kimi K3 at 709 tokens a second
The project put that against 450 tokens a second on Nvidia's GB200, and said all 92 of the model's expert layers run in one kernel.
The project put that against 450 tokens a second on Nvidia's GB200, and said all 92 of the model's expert layers run in one kernel.
Engineers credit a model working in a single Slack channel, and put a fresh web load at 550 milliseconds against 3,085 before the sprint.
Both models are live in the Gemini API and Google AI Studio, and Artificial Analysis puts one of them first on its pronunciation benchmark.
The company says about 950 agents searched sequence databases for 21 hours and turned up a reverse transcriptase next to a repeat array.
Anthropic said on Tuesday that agents running on its Claude models had found a new enzyme system in bacteriophages, which it calls array-associated reverse transcriptases, or ART. The company said the system pairs a reverse transcriptase with an accessory protein of unknown function and a long array of evenly spaced DNA repeats.
Rollouts reads a pull request, writes a monitoring plan, then compares live telemetry against a pre-deploy baseline. It is on Teams and Enterprise plans today.
The directory lists connectors, agents and consultancies from Cursor, CrowdStrike, Harvey and Snowflake, and Anthropic says an existing contract can pay for them.
Voice can now reach a user's email, calendar and Slack, and Plus and Pro subscribers get agentic tasks on mobile.
Custom feeds turn a typed description into a pinned tab on the home page, and reach web and mobile next month.
Cryptologic historian Frode Weierud says a German Army signal from July 1941 gave up its key after the model wrote its own Bombe in two days.
Three researchers say they let language models steer a real car around a cone course, and published the token bill for each attempt.
The beta lets United States merchants check inventory, change prices and edit listings from Anthropic's Claude or Amazon's own Quick assistant.
The model takes text, images and video, and has a context window of about a million tokens. OpenCode has not said who built it.
Nemotron 3 Diarization labels who spoke when for up to 8 speakers, and Nvidia says it cuts diarization error to 14.72% against 19.3% for the next system.
JevBench, published by Benchmark Heaven under an MIT licence, ranks typed decision models on 534 public and 308 sealed decisions, and its authors call the set partial.