News for hackers. Stuff that matters.
Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by bel8 on Saturday October 24, 2026 @10:51AM (#50221840)
Isn't Kimi API 5x cheaper than Opus API?

And it is much more token efficient. There are numbers down in comments of this post.

So most of those points are indeed true. While others look like exaggeration, expected from aggressive marketing campaigns.

Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by kennywinker on Saturday October 24, 2026 @11:06AM (#50221984)
1. Motion design: all opinion

2. 3D tool use: factual, kimi made a thing.

3.a. Open source opus 5.5: lies. Opus 5.5 scores 58 on artificial analysis, kimi 3.1 pro scores 55. So not “literally opus 5.5”.

3.b. 5x cheaper: also seems like lies. kimi 3.1 pro api cost on openrouter: $2 / $10per 1M - opus 5.5 api cost on openrouter: $4 / $20per 1M. Artificial analysis’ cost per task looks about the same, maybe 2.5x cheaper at best.

4. A fact about how it can be run on ~$10k of hardware, and some opinion about how that’s cool.

5. Gap is one month. Potentially factual, but only if 58 == 55 ;)

6. Kimi trained with limited resources: Factual and true afaik.

To be clear, i am a fan of open weight models. I refuse to use anything else. I am just not a fan of lies and astroturfing.

Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by xhrpost on Saturday October 24, 2026 @01:46PM (#50223482)
Are people actually running this on $10k of hardware?
Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by kennywinker on Saturday October 24, 2026 @04:13PM (#50224721)
I don’t know. You can run models that are in the ballpark of this on $1500 of hardware (e.g. strata on 12gb vram + 64gb ram), so I feel like unless you’re training stuff you either run smaller local models, rent your hardware, or pay for tokens.
Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by seunosewa on Sunday October 25, 2026 @04:32PM (#50235775)
Kimi 3.1 is a 2.8T parameter model. You can't run it on a 12GB VRAM GPU. You need 1.4TB of VRAM just to store the parameters.
Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by kennywinker on Sunday October 25, 2026 @05:42PM (#50236336)
Not really. Kimi has 104B active parameters, so afaik that means you’d only need ~64gb vram to run an 8bit quantization of this, and potentially as little as 32gb to run a 4bit quantization. The rest, the MoE experts, can be in system ram… or if you’re willing to take the speed hit, offloaded to ssd. But yes you’d need a lot of system ram… like 1.5TB.

But that’s all kinda moot because I was referring to smaller open models, ones you could easily run on $1500 of hardware.

Re:Kimi K3.1 Pro and K3.1 Pro Highspeed by Forgeties79 on Saturday October 24, 2026 @11:59AM (#50222552)
If you do not make it clear that you are paid by the company that you are writing an opinion about, the truth is borderline immaterial. I automatically do not trust you because you are deliberately obscuring this fact. If it doesn’t influence your opinion, then you shouldn’t have any problem admitting that you are getting paid. This is standard disclosure stuff, a very reasonable expectation.