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

131–140 of 146 posts

Re: Kimi K2.5 Technical Report [pdf]

#131

Post by jmiskovic »

tomaskafka wrote: post #123

It is amazing, but "open source model" means "model I can understand and modify" (= all the training data and processes). Open weights is an equivalent of binary driver blobs everyone hates. "Here is an opaque thing, you have to put it on your computer and trust it, and you can't modify it."

That's unfair. Binary driver blobs are blackmail: "you bought the hardware, but parts of the laptop won't work unless you agree to run this mysterious bundle insecurely". Open weight is more like "here's a frozen brain you can thaw in a safe harness to do your bidding".

Re: Kimi K2.5 Technical Report [pdf]

#132

Post by oefrha »

jgalt212 wrote: post #120
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

What's the point of using an open source model if you're not self-hosting?

Open source models can be hosted by provider, in particular plenty of educational institutions host open source models. You get to choose whatever provider you trust. For instance I used DeepSeek R1 a fair bit last year but never on deepseek.com or through its API.

Re: Kimi K2.5 Technical Report [pdf]

#134

Post by xxr3376 »

zeroxfe wrote: post #20
epolanski wrote: post #13

Earlier quoted context omitted.

You can plug another model in place of Anthropic ones in Claude Code.

That tends to work quite poorly because Claude Code does not use standard completions APIs. I tried it with Kimi, using litellm[proxy], and it failed in too many places.

You can try Kimi's Anthropic-compatible API.

Just connect Claude Code to Kimi's API endpoint and everything works well

https://www.kimi.com/code/docs/en/more/third-party-agents.ht...

Re: Kimi K2.5 Technical Report [pdf]

#135

Post by user_7832 »

orbital-decay wrote: post #108
unleaded wrote: post #83

Seems that K2.5 has lost a lot of the personality from K2 unfortunately, talks in more ChatGPT/Gemini/C-3PO style now. It's not explictly bad, I'm sure most people won't care but it was something that made it unique so it's a shame to see it go. examples to illustrate https://www.kimi.com/share/19c115d6-6402-87d5-8000-000062fec... (K2.5) https://www.kimi.com/share/19c11615-8a92-89cb-8000-000063ee6... (K2)

K2 in your example is using the GPT reply template (tl;dr - terse details - conclusion, with contradictory tendencies), there's nothing unique about it. That's exactly how GPT-5.0 talked. The only model with a strong "personality" vibe was Claude 3 Opus.

> The only model with a strong "personality" vibe was Claude 3 Opus.

Did you have the chance to use 3.5 (or 3.6) Sonnet, and if yes, how did they compare?

As a non-paying user, 3.5 era Claude was absolutely the best LLM I've ever used in terms of having a conversation. It felt like talking to a human and not a bot. Its replies were readable, even if they were several paragraphs long. I've unfortunately never found anything remotely as good.

Re: Kimi K2.5 Technical Report [pdf]

#136

Post by orbital-decay »

user_7832 wrote: post #135
orbital-decay wrote: post #108

Earlier quoted context omitted.

K2 in your example is using the GPT reply template (tl;dr - terse details - conclusion, with contradictory tendencies), there's nothing unique about it. That's exactly how GPT-5.0 talked. The only model with a strong "personality" vibe was Claude 3 Opus.

> The only model with a strong "personality" vibe was Claude 3 Opus. Did you have the chance to use 3.5 (or 3.6) Sonnet, and if yes, how did they compare? As a non-paying user, 3.5 era Claude was absolutely the best LLM I've ever used in terms of having a conversation. It felt like talking to a human and not a bot. Its replies were readable , even if they were several paragraphs long. I've unfortunately never found a…

Pretty poorly in that regard. In 3.5 they killed Claude 3's agency, pretty much reversing their previous training policy in favor of "safety", and tangentially mentioned that they didn't want to make the model too human-like. [1] Claude 3 was the last version of Claude, and one of the very few models in general, that had a character. That doesn't mean it wasn't writing slop though, falling into annoying stereotypes is still unsolved in LLMs.

[1] https://www.anthropic.com/research/claude-character (see the last 2 paragraphs)

Re: Kimi K2.5 Technical Report [pdf]

#139

Post by pstuart »

jgalt212 wrote: post #127
dimava wrote: post #122

Earlier quoted context omitted.

Open source models costs are determined only by electricity usage, as anyone can rent a GPU qnd host them Closed source models cost x10 more just because they can A simple example is Claude Opus, which costs ~1/10 if not less in Claude Code that doesn't have that price multiplier

But Kimi seems so big that renting the necessary number of GPUs is a non trivial exercise.

Exactly! Electricity, hosting, and amortized cost of the GPUs would be the baseline costs.

Re: Kimi K2.5 Technical Report [pdf]

#140

Post by kristianp »

armcat wrote: post #5
zeroxfe wrote: post #2

I've been using this model (as a coding agent) for the past few days, and it's the first time I've felt that an open source model really competes with the big labs. So far it's been able to handle most things I've thrown at it. I'm almost hesitant to say that this is as good as Opus.

Out of curiosity, what kind of specs do you have (GPU / RAM)? I saw the requirements and it's a beyond my budget so I am "stuck" with smaller Qwen coders.

Note that Kimi K2x is natively 4 bit int, which reduces the memory requirements somewhat.
Post reply on HN