Kimi K3 - The world's first open 3T-class model

Kimi K3 is the world's first open 3T-class model — frontier performance across coding, knowledge work, and reasoning, with native multimodality and 1M context.

Add a comment

Replies

Best

Hi everyone!

Isn’t this pretty astonishing?

And the team is still being humble about it:

While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models.

K3 is ’s 2.8T-parameter MoE model with 1M context and native multimodality.

It is built for long-horizon coding and knowledge work, and the results already put it directly into the frontier-model conversation. K3 can work across large repositories, use vision while iterating on games and frontend work, and keep going through much longer agent runs with limited supervision.

You can use it today on , , , or through the .

Full weights arrive on July 27.

And the landscape may look very different after that..!

 This is wild mainly because 1M context + native multimodality + strong coding changes what open model expectations are. A year ago that combo would have sounded like three separate roadmap items.

The full weights on July 27 are the part I’m watching. If teams can actually run or fine tune something this capable without handing everything to a closed API, the frontier model conversation gets a lot more interesting very quickly.

We’re excited to introduce Kimi K3, our new open frontier model built for long-context, multimodal, agentic work!


Kimi K3 has 2.8T parameters, a 1M-token context window, and native multimodal capabilities. Under the hood, it introduces Kimi Delta Attention for much faster decoding in million-token contexts, plus Attention Residuals for more efficient training.

We built K3 especially for long-horizon coding, research, financial analysis, and self-evolving agent workflows — the kinds of tasks where models need to stay coherent, use tools, inspect outputs, recover from failed attempts, and keep going.

Kimi K3 is live now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

Open weights are coming July 27, 2026.

Would love to hear what you build with it.

 so incredibly excited for this! Congratulations on the launch. We can't wait to use it with PromptQL!

put it simply:

"Kimi K3 is the best performing model [for ], ahead of Fable, reaching a comparable success rate in less time."

This is a breakthrough moment for open models. , models were ahead for work requiring correctness and accuracy. Not anymore. LFG!

honestly the multi-file thing blew me away, dropped in like 20 mixed pdfs and slides and it actually pulled everything together coherently

Dropped a 40-page PDF into Kimi and asked it to compare it against a competitor's doc, and it pulled the right figures without me re-uploading anything. The slides maker is rough around the edges but shockingly fast for a first pass.

I'd love to see Kimi integrate with more file types, especially video and audio files. The real-time web search feature sounds incredibly useful, but I'm curious - how does it handle search queries with multiple keywords or complex phrases?

wow, I'm really excited for this new model. Does it also supports image processing?

K3 has a 1M-token context—what's the largest real-world task you've successfully completed end-to-end that simply wasn't practical with previous models?

I wonder why there is a gamboy advance in the slides? Can it make gba rom files from a prompt? Also. whats pricinging going for?

If it can actually produce 3D models like in the demo video, that would be extremely impressive. I've used older versions of Kimi and I've found it pretty impressive, so I'm actually really hyped to try K3 when I get a chance.

12
Next
Last