Newsletter image

Subscribe to the Newsletter

Join 10k+ people to get notified about new posts, news and tips.

Do not worry we don't spam!

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use

Search

GDPR Compliance

We use cookies to ensure you get the best experience on our website. By continuing to use our site, you accept our use of cookies, Privacy Policy, and Terms of Service.

LongCat-2.0

Meituan open-sourced LongCat-2.0: 1.6T parameters, about 48B active, native 1M context, MIT license. It spent two months anonymously topping OpenRouter agent leaderboards as Owl Alpha, and Meituan says it was trained entirely on domestic Chinese chips. Here are the numbers and their asterisks.

Qwen 3.6

Alibaba's Qwen 3.6 ships an open dense 27B that beats the old 397B flagship on coding and fits a 24GB GPU, under Apache 2.0. But the new flagship, Qwen3.6-Max, is closed. What you can download, and how to run it.

MisoTTS

Miso Labs' MisoTTS is an 8B open voice model that clones a speaker from ten seconds of audio, runs on a 24GB GPU, and ships under a license you can build a product on. English-only and unverified on quality, but genuinely open.

Cosmos 3

NVIDIA's Cosmos 3 is an open world model for physical AI: it predicts what happens next in the real world and outputs robot actions, with weights, data, and recipes under a permissive license. What it is, and what you can run.

Kimi K2.7 Code

Moonshot's Kimi K2.7 Code keeps coding quality steady while thinking about 30 percent less, a cost story for long agent runs. The catch: the benchmarks are all first-party. What to believe, and how to run it.

OpenCode

OpenCode is the open-source coding agent for your terminal, model-agnostic, MIT-licensed, and the obvious landing spot after Google shut down Gemini CLI for consumer users. What it is, why it won, and how to point it at open models.

Rio 3.5 Open 397B

Rio de Janeiro's municipal IT company published a 403B open MoE as proof a city could build frontier AI. Then the community checked the weights and found a merge. Here is what Rio 3.5 actually is, and whether you should run it.

The June 2026 Open-Weight Wave: 16 Models in One Week

The first week of June 2026 was the densest stretch of open-weight releases anyone has tracked: frontier models across language, image, audio, video, and 3D, mostly under permissive licenses. What shipped, what to believe, and what it means for builders.

Ideogram 4

Ideogram 4 is a closed image lab's first open weights: a 9.3B from-scratch flow-matching model with the best text rendering of any open model, running on a single 24GB GPU. The catch is the license: free to play with, 300 dollars a month to ship. The honest version.

Nemotron 3 Ultra

NVIDIA's Nemotron 3 Ultra is a 550B hybrid Mamba-MoE built for long-running agents, the most intelligent US open-weights model, and NVIDIA shipped the data and recipes too, under a permissive license. What it is, what is open, and why you still cannot run it at home.

Mistral Large 3

Mistral Large 3 is a 675B Mixture-of-Experts flagship under Apache 2.0, Mistral taking its best model out from behind a research-only license. The licensing is the story; the benchmarks are good, not frontier. Here is the honest version.

GLM-5.2

Z.ai's GLM-5.2 is the first MIT-licensed model you can self-host that beats GPT-5.5 on long-horizon agentic coding, with a real 1M-token context, at about one-sixth the API cost. What changed, the numbers with their asterisks, and how to run it.

ZAYA1-8B

Zyphra released ZAYA1-8B, an 8.4B-total / 760M-active Mixture-of-Experts reasoning model under Apache 2.0. It is the first large MoE pretrained, midtrained, and fine-tuned end-to-end on AMD Instinct MI300X, with no Nvidia in the loop. The single-pass math and code scores are strong, the agentic numbers are not, and every benchmark is Zyphra-reported.