clem 🤗

Cognition

Hacker News

Moonshot AI dropped Kimi K3, and the reception has been unusually loud for an open-weight release.
On Hugging Face, it hit #1 trending with 4,000+ likes faster than any previous model release, according to Clement Delangue. That’s not a small thing — HF trending is a decent proxy for genuine developer interest, not just hype from the vendor’s own Twitter account.
The benchmark claim doing the rounds: Cognition tested it on FrontierCode 1.1 and says K3 is “the first open source model we tested that approaches frontier-level performance” on coding tasks. That’s why it’s already available in Devin Desktop and CLI. Vercel’s Guillermo Rauch went further, calling it “the most powerful open-weight model in the world” — which is the kind of claim that usually ages badly, but he’s backed it with actual infrastructure: K3 is live on Vercel AI Gateway with US-based inference and Zero-Data Retention agreements for anyone who needs them.
The pattern here is interesting. Every major platform — Devin, Vercel, presumably more to follow — is integrating K3 within hours of release. That’s not the usual open-weight story where a model drops, researchers poke at it for a week, and enterprise tooling catches up months later. This one landed with integrations already in place.
What’s still missing: independent evals outside of Cognition’s own benchmark, and any real sense of how it holds up on tasks beyond coding. “Approaches frontier-level” is doing a lot of work in that sentence. The community will stress-test that claim quickly — it always does.

Trends

Moonshot AI released Kimi K3, an open-weight model that hit #1 trending on Hugging Face with 4,000+ likes faster than any previous model. Cognition benchmarked it on FrontierCode 1.1, calling it the first open-source model approaching frontier-level coding performance. It launched with immediate integrations in Devin Desktop, Devin CLI, and Vercel AI Gateway. Notable caveats: the primary benchmark comes from Cognition itself, and independent evals on non-coding tasks are still absent.

Kimi K3 is the open-weight model everyone is actually excited about

Kimi K3 is better than Fable 5 (based on test results) and is not dependent on any country’s government. We’ll test it on real projects.

is there a free way to try it out.?

Is there a cheap way to try it out?

I’ve made some tests and it’s impressive. We talk here in generell about something thats rising over 3 - 4 years now.