Kimi K2 stretches to 256K, Switzerland opens Apertus

What shipped

Moonshot AI

Kimi K2 Instruct 0905

Context goes from 128K to 256K tokens and the agentic coding behavior is tuned, still as open weights.

Swiss AI (EPFL, ETH Zurich, CSCS)

Apertus 70B

A fully open 8B and 70B family from Swiss public institutions, with the training recipe published.

Google

EmbeddingGemma 300M

A small multilingual embedding model built to run on a phone or laptop with no server.

Tencent

HunyuanWorld-Voyager

Takes one image and a camera path and produces a video of a world that stays geometrically consistent.

Tencent

Hunyuan-MT-7B

A 7B translation model covering 33 languages, small enough to self-host.

Anthropic

Anthropic raises Series F

A $13B round at a $183B post-money valuation, with run-rate revenue the company says passed $5B.

OpenAI

Why language models hallucinate

Argues that benchmarks reward confident guessing, so models learn to bluff instead of saying they don't know.

Moonshot shipped a point update this week and I think it's the most useful thing on the list for anyone who builds with agents. Kimi K2 Instruct 0905 doubles the context window to 256K tokens and tunes the model for agentic coding, and it stays open weights.

Why does that matter to someone running a company on AI tooling? Because the long-context, tool-heavy jobs are exactly where I hit limits. A tender pack is hundreds of pages. An estimate pulls from a dozen quotes. When a model that I can host myself holds that much in view, I stop paying per token for the boring middle of the workflow, and I stop worrying about what leaves my machine.

Close behind it is Apertus from Switzerland. EPFL, ETH Zurich and the national supercomputing center released 8B and 70B models, and the point is openness: weights, data recipe and training details. I won't pretend it beats the Chinese labs on raw capability. What it gives you is something a regulated buyer can point at and audit. I've sat in enough compliance meetings to know that is worth more than a leaderboard spot.

The small and the strange

Google put out EmbeddingGemma, a 300M-parameter multilingual embedding model meant for on-device use. Tencent released HunyuanWorld-Voyager, which turns a single image plus a camera path into a video of a consistent 3D world, and Hunyuan-MT-7B, a compact translator for 33 languages. For Arabic and English work, a translator I can run locally is a real tool, not a toy.

Money and a paper

Anthropic closed a $13B Series F at a $183B post-money valuation and said its run-rate revenue passed $5B. I read that less as a valuation story and more as a signal that companies are paying real invoices for agents and coding tools.

OpenAI published a paper on why language models hallucinate. The short version is that current benchmarks score a wrong answer and a blank the same, so guessing wins. That is an incentive problem, and incentives are something an operator can change in his own evals.

What I'd watch next

Whether Qwen answers Kimi on long-context agents. The Qwen team has been teasing something big, and I'd like to see it open.