Reasonable speeds are possible with 4bit quants on 2 512GB Mac Studios (MLX TB4 Ring - see https://x.com/awnihannun/status/1943723599971443134) or even a single socket Epyc system with >1TB of RAM (about the same real world memory throughput as the M Ultra). So $20k-ish to play with it.

For real-world speeds though yeah, you'd need serious hardware. This is more of a "deploy your own stamp" model, less a "local" model.

I tried it a couple of times in comparison to Claude. Kimi wrote much simpler and more readable code than Claude's over-engineered solutions. It missed a few minor subtle edge cases that Claude took care of though.

Claude what? Sonnet? 3.7? 3.5? Opus? 4?

The first question I gave it (a sort of pretty simple recreational math question I asked it to code up for me) and it was outrageously wrong. In fairness, and to my surprise, OpenAI's model also failed with this task, although with some prompting, sort of got it.

Still pretty good, someone with enough resources could distil it down to a more manageable size for the rest of us.

I asked it to give me its opinion on a mail I'm writing. 95% of its content is quotes from famous authors, and the 5% I wrote is actually minimal glue in-between.

All the models I tested, which includes Sonnet 4, DeepSeekR1, 4o and Gemini 2.5 understand this isn't your normal email and what I ask is literary/philosophical criticism, not remarks about conventions, formatting or how to convey my message in a more impactful way.

    Quick fix:
    - Replace half the commas with periods.
    - Let one metaphor bloom and then stop.
    - Read it aloud; wherever you inhale, insert a line break.
    
    Before:
    >And if the collective imagination is so seized by it, it's because
    we are already—not in some other world—but in this very life, in a
    state of socio-, photo-, and videosynthesis.
    
    After:
    >The collective imagination is seized.
    >We are not in another world.
    >We are in this life, in a state of videosynthesis.

Yes, this quote is by Baudrillard. None of the other models fixated on the fact that it’s an email (I only used the word once in the prompt). My gut feeling is that this reflects not so much a lack of intelligence as a difference in model personality. Here's what it replied when I shared Gemini's analysis. The point was to have it understand that:

    Gemini, second sentence: It is less a simple email than an intellectual
    and personal exploration still in the making.

Point deflected as soon as understood:

    **Gemini is polite and pedagogical; I’m going to be blunt.**
    
    1. **This isn’t an email, it’s a trip.**
       ...
    
    2. **The content is good, the format is not.**
       ...
    
    3. **Gemini is right about one thing:**
       ...
    
       * cut 80% of the material,
       * ask **one clear question**
       * put the rest in a separate Google Doc or PDF for posterity.
    
    4. **Brutal summary:**
       Keep your draft for yourself — it’s valuable. But for the email, write a new one in **three sentences**:
    
       * thank them,
       * recall the context in one line,
       * ask the question.
    
    You’ve already got Gemini’s answer to stroke your ego — now give me the short version you’re actually going to send.

The solution to sycophancy is not disparagement (misplaced criticism). The classical true/false positive/negative dilemma is at play here. I guess the bot got caught in the crossfire of 1°) its no-bullshit attitude (it can only be an attitude) 2°) preference for delivering blunt criticism over insincere flattery 3°) being a helpful assistant. Remove point 3°), and it could have replied: "I'm not engaging in this nonsense". Preserve it and it will politely suggest that you condense your bullshit text, because shorter explanations are better than long winding rants (it's probably in the prompt).

For what it's worth, I think Kimi's modified MIT license still meets the OSI definition of "open source." For example, the explicitly OSI-approved "Attribute Assurance License"[1] contains similar wording:

> each time the resulting executable program or a program dependent thereon is launched, a prominent display (e.g., splash screen or banner text) of the Author’s attribution information

[1] https://opensource.org/license/attribution-php

At this point, they have to be training it. At what point will you start using something else?

I'm glad we are looking to build nuclear reactors so we can do more of this...

Much better than that of Grok 4.

That's perhaps the best one I've seen yet! For an open weight model, this performance is of course particularly remarkable and impactful.

wow!

>If someone took Kimi k2 and trained a reasoning model with it

I imagine that's what they are going at MoonshotAI right now

Why hasn’t Kimis current and older models been benchmarked and added to Artificial analysis yet?

I have a different experience in chatting/creative writing. It tends to overuse certain speech patterns without repeating them verbatim, and is strikingly close to the original R1 writing, without being "chaotic" like R1 - unexpected and overly dramatic sci-fi and horror story turns, "somewhere, X happens" at the end etc.

Interestingly enough, EQ-Bench/Creative Writing Bench doesn't spot this despite clearly having it in their samples. This makes me trust it even less.

Paired with programming tools like Claude Code, it could be a low-cost/open-source replacement for Sonnet

Reasonable question, simple answer: "New breed of model" is overstating it — all these models for years have been fine-tuned using reinforcement learning on a variety of tasks, it's just that the set of tasks (and maybe the amount of RL) has changed over time to include more tool use tasks, and this has made them much, much better at the latter. The explosion of tools like Claude Code this year is driven by the models just being more effective at it. The orchestration external to the model you mention is what people did before this year and it did not work as well.

"Agentic" and "agent" can mean pretty much anything, there are a ton of different definitions out there.

When an LLM says it's "agentic" it usually means that it's been optimized for tool use. Pretty much all the big models (and most of the small ones) are designed for tool use these days, it's an incredibly valuable feature for a model to offer.

I don't think this new model is any more "agentic" than o3, o4-mini, Gemini 2.5 or Claude 4. All of those models are trained for tools, all of them are very competent at running tool calls in a loop to try to achieve a goal they have been given.

It is not a silly question. The various flavors of LLM have issues with reliability. In software we expect five 9s, LLMs aren't even a one 9. Early on it was reliability of them writing JSON output. Then instruction following. Then tool use. Now it's "computer use" and orchestration.

Creating models for this specific problem domain will have a better chance at reliability, which is not a solved problem.

Jules is the gemini coder that links to github. Half the time it doesn't create a pull request and forgets and assumes I'll do some testing or something. It's wild.

I'm new too. Found this article helpful: https://crawshaw.io/blog/programming-with-agents

> I'm sorry if my questions sound silly. Following AI world is like following JavaScript world.

You are more right than you could possibly imagine.

TL;DR: "agentic" just means "can call tools it's been given access to, autonomously, and then access the output" combined with an infinite loop in which the model runs over and over (compared to a one-off interaction like you'd see in ChatGPT). MCP is essentially one of the methods to expose the tools to the model.

Is this something the models could do for a long while with a wrapper? Yup. "Agentic" is the current term for it, that's all. There's some hype around "agentic AI" that's unwarranted, but part of the reason for the hype is that models have become better at tool calling and using data in their context since the early days.

Someone at openai did say it was too big to host at home, so you could be right. They will probably be benchmaxxing, right now, searching for a few evals they can beat.

According to the benchmarks, Kimi K2 beats GPT-4.1 in many ways. So to "compete", OpenAI would have to release the GPT-4.1 weights, or a similar model. Which, I guess, they likely won't do.

To me K2 is the Kotlin 2.0 compiler. https://blog.jetbrains.com/kotlin/2023/02/k2-kotlin-2-0/