Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

kimi.com

Post reply on HN
Print view View original

91–100 of 251 posts

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#91

Post by ffsm8 »

Barathkanna wrote: post #60
embedding-shape wrote: post #52

Earlier quoted context omitted.

I'd love to see the prompt processing speed difference between 16× H100 and 2× Mac Studio.

I asked GPT for a rough estimate to benchmark prompt prefill on an 8,192 token input. • 16× H100: 8,192 / (20k to 80k tokens/sec) ≈ 0.10 to 0.41s • 2× Mac Studio (M3 Max): 8,192 / (150 to 700 tokens/sec) ≈ 12 to 55s These are order-of-magnitude numbers, but the takeaway is that multi H100 boxes are plausibly ~100× faster than workstation Macs for this class of model, especially for long-context prefill.

You do realize that's entirely made up, right?

Could be true, could be fake - the only thing we can be sure of is that it's made up with no basis in reality.

This is not how you use llms effectively, that's how you give everyone that's using them a bad name from association

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#92

Post by whizzter »

motoboi wrote: post #80
bertili wrote: post #34

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".

What amazes me is why would someone spend millions to train this model and give it away for free. What is the business here?

Chinese state that maybe sees open collaboration as the way to nullify any US lead in the field, concurrently if the next "search-winner" is built upon their model the Chinese worldview that Taiwan belongs to China and Tiamen Square massacre never happened.

Also their license says that if you have a big product you need to promote them, remember how Google "gave away" site searche widgets and that was perhaps one of the major ways they gained recognition for being the search leader.

OpenAI/NVidia is the Pets.com/Sun of our generation, insane valuations, stupid spend, expensive options, expensive hardware and so on.

Sun hardware bought for 50k USD to run websites in 2000 are less capable than perhaps 5 dollar/month VPS's today?

"Scaling to AGI/ASI" was always a fools errand, best case OpenAI should've squirreled away money to have a solid engineering department that could focus on algorithmic innovations but considering that Antrophic, Google and Chinese firms have caught up or surpassed them it seems they didn't.

Once things blows up, those closed options that had somewhat sane/solid model research that handles things better will be left and a ton of new competitors running modern/cheaper hardware and just using models are building blocks.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#93

Post by Alifatisk »

Reubend wrote: post #13

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.

Yup, I experience the same. I don't know what they do to achieve this but it gives them this edge, really curious to learn more about what makes it so good at it.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#94

Post by MattRix »

throwaw12 wrote: post #67

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.

Opus 4.5 only came out two months ago, and yes Anthropic spends a lot of effort making it particularly good at coding.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#95

Post by ggdG »

motoboi wrote: post #80
bertili wrote: post #34

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".

What amazes me is why would someone spend millions to train this model and give it away for free. What is the business here?

I think this fits into some "Commoditize The Complement" strategy.

https://gwern.net/complement

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#96

Post by Alifatisk »

mrklol wrote: post #19
oneneptune wrote: post #7

Earlier quoted context omitted.

I'm curious about the "cheaper" claim -- I checked Kimi pricing, and it's a $200/mo subscription too?

They also have a $20 and $40 tier.

If you bargain with their bot Kimmmmy (not joking), you can even get lower pricing.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#97

Post by zozbot234 »

XCSme wrote: post #89
rvnx wrote: post #38

Earlier quoted context omitted.

You have a team lead that establishes a list of tasks that are needed to achieve your mission then it creates a list of employees, each of them is specialized for a task, and they work in parallel. Essentially hiring a team of people who get specialized on one problem. Do one thing and do it well.

But in the end, isn't this the same idea with the MoE? Where we have more specialized "jobs", which the model is actually trained for. I think the main difference with agents swarm is the ability to run them in parallel. I don't see how this adds much compared to simply sending multiple API calls in parallel with your desired tasks. I guess the only difference is that you let the AI decide how to split those requests…

Nope. MoE is strictly about model parameter sparsity. Agents are about running multiple small-scale tasks in parallel and aggregating the results for further processing - it saves a lot of context length compared to having it all in a single session, and context length has quadratic compute overhead so this matters. You can have both.

One positive side effect of this is that if subagent tasks can be dispatched to cheaper and more efficient edge-inference hardware that can be deployed at scale (think nVidia Jetsons or even Apple Macs or AMD APU's) even though it might be highly limited in what can fit on the single node, then complex coding tasks ultimately become a lot cheaper per token than generic chat.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#98

Post by mzl »

XCSme wrote: post #87
jumploops wrote: post #4

> 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!

> Kimi K2.5 can self-direct an agent swarm Is this within the model? Or within the IDE/service that runs the model? Because tool calling is mostly just the agent outputting "call tool X", and the IDE does it and returns the data back to AI's context

An LLM model only outputs tokens, so this could be seen as an extension of tool calling where it has trained on the knowledge and use-cases for "tool-calling" itself as a sub-agent.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#99

Post by Balinares »

motoboi wrote: post #80
bertili wrote: post #34

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".

What amazes me is why would someone spend millions to train this model and give it away for free. What is the business here?

Speculating: there are two connected businesses here, creating the models, and serving the models. Outside of a few moneyed outliers, no one is going to run this at home. So at worst opening this model allows mid-sized competitors to serve it to customers from their own infra -- which helps Kimi gain mindshare, particularly against the large incumbents who are definitely not going to be serving Kimi and so don't benefit from its openness.

Given the shallowness of moats in the LLM market, optimizing for mindshare would not be the worst move.

Re: Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model

#100

Post by Balinares »

throwaw12 wrote: post #67

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.

I replaced Opus with Gemini Pro and it's just plain a better coder IMO. It'll restructure code to enable support for new requirements where Opus seems to just pile on more indirection layers by default, when it doesn't outright hardcode special cases inside existing functions, or drop the cases it's failing to support from the requirements while smugly informing you you don't need that anyway.
Post reply on HN