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

Kimi K2.5 Technical Report [pdf]

github.com

Post reply on HN
Print view View original

21–30 of 146 posts

Re: Kimi K2.5 Technical Report [pdf]

#21

Post by KolmogorovComp »

explorigin wrote: post #8
thesurlydev wrote: post #3

Earlier quoted context omitted.

Can you share how you're running it?

https://unsloth.ai/docs/models/kimi-k2.5 Requirements are listed.

To save everyone a click

> The 1.8-bit (UD-TQ1_0) quant will run on a single 24GB GPU if you offload all MoE layers to system RAM (or a fast SSD). With ~256GB RAM, expect ~10 tokens/s. The full Kimi K2.5 model is 630GB and typically requires at least 4× H200 GPUs. If the model fits, you will get >40 tokens/s when using a B200. To run the model in near full precision, you can use the 4-bit or 5-bit quants. You can use any higher just to be safe. For strong performance, aim for >240GB of unified memory (or combined RAM+VRAM) to reach 10+ tokens/s. If you’re below that, it'll work but speed will drop (llama.cpp can still run via mmap/disk offload) and may fall from ~10 tokens/s to <2 token/s. We recommend UD-Q2_K_XL (375GB) as a good size/quality balance. Best rule of thumb: RAM+VRAM ≈ the quant size; otherwise it’ll still work, just slower due to offloading.

Re: Kimi K2.5 Technical Report [pdf]

#22

Post by naragon »

margorczynski wrote: post #6

I wonder how K2.5 + OpenCode compares to Opus with CC. If it is close I would let go of my subscription, as probably a lot of people.

I've been using K2.5 with OpenCode to do code assessments/fixes and Opus 4.5 with CC to check the work, and so far so good. Very impressed with it so far, but I don't feel comfortable canceling my Claude subscription just yet. Haven't tried it on large feature implementations.

Re: Kimi K2.5 Technical Report [pdf]

#23

Post by zeroxfe »

derac wrote: post #9

I really like the agent swarm thing, is it possible to use that functionality with OpenCode or is that a Kimi CLI specific thing? Does the agent need to be aware of the capability?

It seems to work with OpenCode, but I can't tell exactly what's going on -- I was super impressed when OpenCode presented me with a UI to switch the view between different sub-agents. I don't know if OpenCode is aware of the capability, or the model is really good at telling the harness how to spawn sub-agents or execute parallel tool calls.

Re: Kimi K2.5 Technical Report [pdf]

#25

Post by zeroxfe »

BeetleB wrote: post #18
zeroxfe wrote: post #14

Earlier quoted context omitted.

I'm not running it locally (it's gigantic!) I'm using the API at https://platform.moonshot.ai

Just curious - how does it compare to GLM 4.7? Ever since they gave the $28/year deal, I've been using it for personal projects and am very happy with it (via opencode). https://z.ai/subscribe

It's waaay better than GLM 4.7 (which was the open model I was using earlier)! Kimi was able to quickly and smoothly finish some very complex tasks that GLM completely choked at.

Re: Kimi K2.5 Technical Report [pdf]

#26

Post by chillacy »

alansaber wrote: post #16
behnamoh wrote: post #10

It's a decent model but works best with kimi CLI, not CC or others.

Why do you think that is?

I heard it's because the labs fine tune their models for their own harness. Same reason why claude does better in claude code than cursor.

Re: Kimi K2.5 Technical Report [pdf]

#27

Post by eknkc »

margorczynski wrote: post #6

I wonder how K2.5 + OpenCode compares to Opus with CC. If it is close I would let go of my subscription, as probably a lot of people.

It is not opus. It is good, works really fast and suprisingly through about its decisions. However I've seen it hallucinate things.

Just today I asked for a code review and it flagged a method that can be `static`. The problem is it was already static. That kind of stuff never happens with Opus 4.5 as far as I can tell.

Also, in an opencode Plan mode (read only). It generated a plan and instead of presenting it and stopping, decided to implement it. Could not use the edit and write tools because the harness was in read only mode. But it had bash and started using bash to edit stuff. Wouldn't just fucking stop even though the error messages it received from opencode stated why. Its plan and the resulting code was ok so I let it go crazy though...

Re: Kimi K2.5 Technical Report [pdf]

#28

Post by Imustaskforhelp »

eknkc wrote: post #11
thesurlydev wrote: post #3

Earlier quoted context omitted.

Can you share how you're running it?

I've been using it with opencode. You can either use your kimi code subscription (flat fee), moonshot.ai api key (per token) or openrouter to access it. OpenCode works beautifully with the model. Edit: as a side note, I only installed opencode to try this model and I gotta say it is pretty good. Did not think it'd be as good as claude code but its just fine. Been using it with codex too.

I tried to use opencode for kimi k2.5 too but recently they changed their pricing from 200 tool requests/5 hour to token based pricing.

I can only speak from the tool request based but for some reason anecdotally opencode took like 10 requests in like 3-4 minutes where Kimi cli took 2-3

So I personally like/stick with the kimi cli for kimi coding. I haven't tested it out again with OpenAI with teh new token based pricing but I do think that opencode might add more token issue.

Kimi Cli's pretty good too imo. You should check it out!

https://github.com/MoonshotAI/kimi-cli

Re: Kimi K2.5 Technical Report [pdf]

#30

Post by Gracana »

KolmogorovComp wrote: post #21
explorigin wrote: post #8

Earlier quoted context omitted.

https://unsloth.ai/docs/models/kimi-k2.5 Requirements are listed.

To save everyone a click > The 1.8-bit (UD-TQ1_0) quant will run on a single 24GB GPU if you offload all MoE layers to system RAM (or a fast SSD). With ~256GB RAM, expect ~10 tokens/s. The full Kimi K2.5 model is 630GB and typically requires at least 4× H200 GPUs. If the model fits, you will get >40 tokens/s when using a B200. To run the model in near full precision, you can use the 4-bit or 5-bit quants. You can use…

I'm running the Q4_K_M quant on a xeon with 7x A4000s and I'm getting about 8 tok/s with small context (16k). I need to do more tuning, I think I can get more out of it, but it's never gonna be fast on this suboptimal machine.
Post reply on HN