Claim◆Video · 24:23 — 25:53

Kimmy K3 is the first open-weight AI model whose benchmarks put it on par with frontier closed models like Opus and GPT.

CJ reports that Kimmy K3, a new 2.8-trillion-parameter open model released by a Chinese lab, is the first open model to match frontier proprietary models like Opus and GPT-5.5 on coding benchmarks. ✦ AI generated

CJ Reynolds · Syntax · 2026-07-20 · original ↗

Share card ↓Export .md Export .txt

starts at this moment · 24:23

Elicited by

“You're going to give us a rundown of what? Inkling, Quinn, Kimmy, View, what else?”

up until now most of the open models have been decent but have never really been able to be on par with like Opus level models or GPT55 level models. But now with Kimmy, we're there and the benchmarks look good and the rumblings of people using it on Twitter look good as well.

verbatim transcript · starts at 24:23

Transcript · around this moment

24:23really been able to be on par with like Opus level models or GPT55 level models. But now with Kimmy, we're there and the benchmarks look good and the rumblings of people using it on Twitter look good as well. Uh but this is a 2.8 8 trillion parameter model. It has a million token context window. Uh, and it's natively multimodal, meaning you can pass it images, PDFs, documents, and it can

24:49figure it can basically figure out what's in them and respond accordingly just like you're used to when you're talking on chatgbt.com. It can do all of those things. Um, and the the thing to look at are these these coding coding benchmarks. Um now again benchmarks are just benchmarks but >> but if you look at these charts Kimmy K3 is sitting for a deep SW. So that's one

25:17of the benchmarks. It's sitting above GPT 5.5. It's like.5% better and just below Fable 5. Um and a few points below GPD 5.6 Soul. >> Wow. >> Yeah. And so on all of these benchmarks, it's either it's like in the top three and that's comparing it to Opus and Fable and GPT56. So the fact that you can eventually run this on your own, they're saying that open weights will be released by July

25:4627th. So right now, if you want to use this model, you do have to talk to Kimmy's servers, which are hosted in China. So um don't do it on anything you wouldn't want them to have your source code of. Uh, but on July 27th, there'll be an open version, so you can run it on GPUs in the cloud. This is a massive model. This is not something you're

26:05going to be able to run on like a Mac Mini or anything like that. You need lots of GPUs. Um, but once they release the open weights, we're going to be seeing probably a lot of people uh making this model available. And uh, for instance, um, GitHub Copilot made Kimmy K2.5 available via their API. And so eventually you're probably going to get access to Kimmy K3 as well. Uh so yeah,

✳ GPT-5.5 ✳ Kimmy K3 ✳ Opus

Around this claim

Mechanism · 1

Chinese open-weight models like Kimi K3 are genuinely catching up to Western frontier models in performance, not just PR, as evidenced by demand overwhelming Kimi's ability to sign up new users.Jason · 20VC · conf 80%

In practice · 1

Open-weight models can be run locally, fine-tuned with proprietary data, and used entirely offline — this is a real benefit to US companies and breeds innovation independent of frontier labs.CJ · Syntax · conf 85%

Context · 1

"Open weight" only means a model's trained numerical parameters are published; it does not disclose the training process or data and is a separate concept from open-source licensing.CJ Reynolds · Syntax · conf 60%

This moment responds to

gives example → Kimi K3 shows open weight models can stand toe-to-toe with proprietary frontier ones.Simon Willison · Simon Willison's Weblog gives example → Open-weight models are not the same as open source: their weights are visible but the training data and creation process remain hidden, and their licenses vary dramatically from permissive (MIT for DeepSeek) to restrictive (Meta Llama's caps and competitor restrictions).CJ · Syntax