47 comments

[ 3.8 ms ] story [ 69.0 ms ] thread
> For complex tasks, Kimi K2.5 can self-direct an agent swarm with up to 100 sub-agents, executing parallel workflows across up to 1,500 tool calls.

> K2.5 Agent Swarm improves performance on complex tasks through parallel, specialized execution [..] leads to an 80% reduction in end-to-end runtime

Not just RL on tool calling, but RL on agent orchestration, neat!

1,500 tool calls per task sounds like a nightmare for unit economics though. I've been optimizing my own agent workflows and even a few dozen steps makes it hard to keep margins positive, so I'm not sure how this is viable for anyone not burning VC cash.
Those are some impressive benchmark results. I wonder how well it does in real life.

Maybe we can get away with something cheaper than Claude for coding.

Kimi was already one of the best writing models. Excited to try this one out
Huggingface Link: https://huggingface.co/moonshotai/Kimi-K2.5

1T parameters, 32b active parameters.

License: MIT with the following modification:

Our only modification part is that, if the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display "Kimi K2.5" on the user interface of such product or service.

Cursor devs, who go out of their way to not mention their Composer model is based on GLM, are not going to like that.
Actually open source, or yet another public model, which is the equivalent of a binary?

URL is down so cannot tell.

I've read several people say that Kimi K2 has a better "emotional intelligence" than other models. I'll be interested to see whether K2.5 continues or even improves on that.
I'll test it out on mafia-arena.com once it is available on Open Router
I love the Kimi response style. It's much more concise, without all the unnecessary "great question!"s and other annoying AI stuff
There are so many models, is there any website with list of all of them and comparison of performance on different tasks?
Curious what would be the most minimal reasonable hardware one would need to deploy this locally?
The chefs at Moonshot have cooked once again.
As your local vision nut, their claims about "SOTA" vision are absolutely BS in my tests.

Sure it's SOTA at standard vision benchmarks. But on tasks that require proper image understanding, see for example BabyVision[0] it appears very much lacking compared to Gemini 3 Pro.

[0] https://arxiv.org/html/2601.06521v1

Gemini remains the only usable vision fm :(
K2 0905 and K2 Thinking shortly after that have done impressively well in my personal use cases and was severely slept on. Faster, more accurate, less expensive, more flexible in terms of hosting and available months before Gemini 3 Flash, I really struggle to understand why Flash got such positive attention at launch.

Interested in the dedicated Agent and Agent Swarm releases, especially in how that could affect third party hosting of the models.

The "Deepseek moment" is just one year ago today!

Coincidence or not, let's just marvel for a second over this amount of magic/technology that's being given away for free... and how liberating and different this is than OpenAI and others that were closed to "protect us all".

There's been so many moments that folks not really heavy into LLM have missed, DeepSeekR1 was great, but so was all the "incremental" improvements, v3-0324, v3.1, v3.1-terminus, and now v3.2-speciale. With that this is the 3rd great Kimi model, then GLM has been awesome, since 4.5, with 4.5, 4.5-air, 4.6, 4.7 and now 4.7 flash. Minimax-M2 has also been making waves lately. ... and i'm just talking about the Chinese model without adding the 10+ Qwen models. Outside of Chinese models, mistral-small/devstral, gemma-27b-it, gpt-oss-120b, seed-os have been great, and I'm still talking about just LLM, not image, audio or special domain models like deepseek-prover and deepseek-math. It's really a marvel what we have at home. I cancelled OpenAI and Anthropic subscription 2 years ago once they started calling for regulation of open models and I haven't missed them one bit.
What's your hardware/software setup?
I don't get this "agent swarm" concept. You set up a task and they boot up 100 LLMs to try to do it in parallel, and then one "LLM judge" puts it all together? Is there anywhere I can read more about it?
Is this actually good or just optimized heavily for benchmarks? I am hopefully its the former based on the writeup but need to put it through its paces.
A realistic setup for this would be a 16× H100 80GB with NVLink. That comfortably handles the active 32B experts plus KV cache without extreme quantization. Cost-wise we are looking at roughly $500k–$700k upfront or $40–60/hr on-demand, which makes it clear this model is aimed at serious infra teams, not casual single-GPU deployments. I’m curious how API providers will price tokens on top of that hardware reality.
About 600GB needed for weights alone, so on AWS you need an p5.48xlarge (8× H100) which costs $55/hour.
Have you all noted that the latest releases (Qwen3 max thinking, now Kimi k2.5) from Chinese companies are benching against Claude opus now and not Sonnet? They are truly catching up, almost at the same pace?
They are, in benchmarks. In practice Anthropic's models are ahead of where their benchmarks suggest.
Pretty cute pelican
Congratulations, great work Kimi team.

Why is that Claude still at the top in coding, are they heavily focused on training for coding or is it their general training is so good that it performs well in coding?

Someone please beat the Opus 4.5 in coding, I want to replace it.

Gemini 3 pro is way better than Opus especially for large codebases.
Glad to to see open source models are catching up and treat vision as first-class citizen (a.k.a native multimodal agentic model). GLM and Qwen models takes different approach, by having a base model and a vision variant (glm-4.6 vs glm-4.6v).

I guess after Kimi K2.5, other vendors are going to the same route?

Can't wait to see how this model performs on computer automation use cases like VITA AI Coworker.

https://www.vita-ai.net/