Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
Can anyone explain the whole KDS - Kimi Derangement Syndrome that is going on? I kinda got the hype around deepseek - something that cheap to be that good. But Kimi i don't understand. It is quite expensive and it seems that no one is doing something interesting with open weights models once they are released from the labs. So what gives?
Some of it's just reaction to the marketing talk: there's a lot of people who haven't run Kimi, or only have run it through OpenRouter, who treat it as Fable equivalent because the benchmarks were good, and a Fable version that can't 'go away' is politically meaningful for the broader open-source movement. Whether they'll actually use it is a separate question.
It does also somewhat reduce the maximum cope levels for anti-AI groups. There's a theory that the 'real' cost of frontier model inference is being subsidized by orders of magnitude, and having the weights makes the upper bound of actual inference costs limited.
A lot of the open model uses aren't something you'll hear about in public. The strongest arguments are for rolling your own inference server are those environments where you can't publish anything, after all. And even if they're only a few steps behind, a lot of 'success' from an open-model is natively less interesting than the 'original' accomplishment from a frontier one, simply because if someone using Kimi K3 to disprove the Jacobian Conjecture literally today it's could plausibly not be independently discovered. Lower-profile stuff like the ThinkingMachines Inkling model put this on the table, but Kimi is at least close to frontier capabilities in ways that Inkling just isn't.
I think the more interesting stuff is happening with the lighter-weight models, stuff in the 100B-250B parameter range. But there's more value than just the dollar/outcome metric.
What I meant is that when llama dropped people went crazy with it. It spawned a whole ecosystem. Now china releases amazing open source models and it is like shouting in the void.
Ah. For Kimi K3 specifically, the weights haven't even been released yet, so it's basically just a (paid) preview mode. The mega-parameter models are also pretty much impossible to tune as an individual or small organization, and even the 200B-1T range is expensive. Model merges in this range mostly haven't worked, to the limited extent they worked cross-family to start with.
Most of the community ecosystems are focused more on the <50B range. Some of the lower visibility is just fragmentation, or the base models being good enough for a lot of purposes, or technical limitations (MoE models are much nicer to low-VRAM users, and also much harder to finetune), but there's just some awkward side effects of the highest-profile models being way outside of home enthusiast use. That said, Qwen3.6 got a massive amount of attention for that lower-range, and I'd expect any 3.8 <100B model gets similar focus.
And a lot of major providers have kinda settled on things as 'good enough'. NovelAI's highest-end model is a GLM-4.6 finetune at 355B param, which is great and all until you look at the release dates.
The movement toward tool use also makes things weird. There's no "here's an opencode project" sharing environment like there is Huggingface or CivitAI, for both obvious reasons (the LLMs can build it themselves... if you know enough to ask them the right questions) and technical ones (opencode has terrible project design and doesn't really distinguish templates from content as an intentional choice). But they do genuinely offer really strong ways to customize an LLM to your use case. So that's probably drawn a bit of the focus away from some of the finetunes.
More options
Context Copy link
They're too big and expensive to do fun things with. You could quantise or frankenmerge or finetune the llama 7b models at home, and get features and performance that the big American companies wouldn't give you because it made the lawyers and wokes and little old ladies sad. Good luck trying that with Kimi K3.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link