Moonshot AI releases Kimi K3, a 2.8T-parameter AI model that it says rivals Opus 4.8 and GPT-5.5, and plans to release model weights by July 27

Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals …

Kimi 2026-07-16

Context & Ripple Effects

Moonshot’s Kimi line has moved from K2’s large mixture-of-experts architecture to open-weight K2.6 and subsequent coding-focused updates, with the company emphasizing long-horizon and agentic workloads. K3 is the next scale-up in that progression.

Related coverage had already signaled a 2T–3T-parameter K3 and positioned it against leading US models. The announced weight release makes the competitive claim consequential beyond a hosted-model launch.

First-order effects

Moonshot gains a new flagship model to market against Opus 4.8 and GPT-5.5, while its planned weight release gives developers a route to evaluate and deploy K3 outside a purely proprietary API offering.
The release raises the practical stakes of Moonshot’s claimed frontier performance: users and researchers will be able to test whether the model’s capabilities hold across coding, reasoning, and agentic tasks.

Second-order effects

Competing model providers face added pressure to differentiate through verified performance, inference efficiency, tooling, licensing, or distribution rather than benchmark claims alone.
Open-model developers and enterprises can compare a much larger Moonshot model with K2.6 and other available weights, potentially shifting experimentation toward models that can be adapted or self-hosted.

Third-order effects

If frontier-scale weights are released more often, the boundary between closed frontier labs and the open-model ecosystem could narrow, moving more competition to deployment cost, customization, and developer infrastructure.
The K3 launch is another test of the reported narrowing US–China frontier-model gap; its longer-term significance depends on independent evaluations and on whether release terms permit broad commercial use.

The trend: Frontier AI competition is expanding from closed-model capability races into a contest over whether very large, high-performing models can be distributed as weights as well as services.

Discussion

@zijing_wu Zijing Wu on x
Chinese AI start-up Moonshot to launch model challenging Anthropic's lead * Set to release as early as tonight * 2-3T, largest Chinese model to date * Benchmark performance Opus 4.8 < K3 < Fable * Attention Residuals & Kimi Linear * Fundraising at $31.5B https://as.ft.com/...
@sriramk Sriram Krishnan on x
kimi k3 is a big moment with multiple implications for the entire industry.
@emostaque Emad on x
something something recursive self improvement [image]
@scaling01 @scaling01 on x
The moonshot was successful the small chinese lab is taking down trillion dollar companies
@aravsrinivas Aravind Srinivas on x
Incredible
@kimi_moonshot @kimi_moonshot on x
Internal knowledge work bench Beyond public benchmarks, Kimi K3 Max also shows consistent gains on our internal benchmarks, which are built from recurring patterns and challenges in real-world user-agent workflows. It scores 75.5 on Online Exp Bench, 73.5 on DECK-Bench, and [imag…
@kimi_moonshot @kimi_moonshot on x
K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural updates designed to improve how information flows across sequence length and model depth. We have also scaled up Mixture of Experts (MoE) sparsity, effectively activating 16 out of [ima…
@kimi_moonshot @kimi_moonshot on x
Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional [ima…
@natolambert Nathan Lambert on x
@arena @Kimi_Moonshot at this point the distillation arguments need to die and understand that China is also very good at building models
@emostaque Emad on x
Kimi K3 by @Kimi_Moonshot is a true frontier model, open source! Congrats! We don't have all details (training tokens etc), but would estimate fp16 pretraining w/ muon => MXFP4/MXFP8 for SFT stage ~1e25 flops total (~Inkling!) Total cost: $15-$25m Scale isn't all you need! [image…
@tszzl Roon on x
the era of the chinese labs being far behind is over, Kimi is at least on par with the modern public frontier models. people have to think differently now without any competitive margin built in
@scaling01 @scaling01 on x
Kimi-K3 is 10th in the Text Arena [image]
@yacinemtb Kache on x
insane. they distilled american frontier models that haven't even been invented yet. damn chinese
@scaling01 @scaling01 on x
what's pretty clear, even without knowing all of these other benchmarks is that China now has models that are genuinely useful for speeding up their own model development
@cline @cline on x
Moonshot's new Kimi K3 is the largest open weight model ever at 2.8T param, and scores the same as GPT-5.6 and Fable 5 on benchmarks. However, it is signficantly cheaper at $3/M input, $15/M output - same price as Sonnet 5. This is a game changing milestone for open weights. [ima…
@arena @arena on x
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, [i…
@emostaque Emad on x
US labs gonna end up distilling Chinese models
@kimi_moonshot @kimi_moonshot on x
Meet Kimi K3 [video]
@_maxblade Max Blade on x
KIMI K3 is here, and it blew my mind 🤯 yes, it is expensive and very slow Took over 30 minutes, and blew an ENTIRE $20 kimi sub before it even finished the one prompt for this game. BUT OMG... the output geniuinly has fable 5 level magic built into it, this model falls [video]
@mweinbach Max Weinbach on x
Ok it's up TLDR Kimi K3 is fine but there are cheaper and better models Muse Spark 1.1 and Kimi K2.6 should not be considered at all. https://csbench.com/...
@zephyr_z9 @zephyr_z9 on x
Crazy big boi numbers from Kimi K3 [image]
@dhh @dhh on x
I just hoped back on Kimi K2.7 for some work yesterday. Hadn't been using it since GPT 5.6 and Fable. But great reminder of just how good these open-weight models have gotten. For what I was doing, Kimi was faster and just as good. Excited for K3!
@yuhasbeentaken @yuhasbeentaken on x
They distilled Fable, now they're distilling Anthropic's marketing. That is a joke about Kimi K3. But,it's working. Copy the frontier, undercut the price, ship faster than the lab you copied from. If K3 lands anywhere near the rumors, expect this to keep happening. Not because [i…
@emollick Ethan Mollick on x
Kimi K3 seems really good, closest to the frontier yet, but also wow does the model/harness love to loop back over and over again over tasks tweaking and changing things at max level. Anyhow, examples incoming, when they finish.
@kimmonismus @kimmonismus on x
Kimi k3 is being released tonight, via FT -2-3t parameters (Opus4.8 has about 1.5t) -1m context -Expected to exceed Opus 4.8 performance! The time when China was six months behind is over. History is presumably being made today. [image]
@yuchenj_uw Yuchen Jin on x
> Kimi K3 ranks only behind Claude Fable 5 Max and GPT-5.6 Sol Max, and surpasses Claude Opus 4.8 Max's score of 1600. I'm super excited for Kimi K3 to be open-sourced! [image]
@jun_song Jun Song on x
Kimi-K3 vs Opus-4.8 Flappybird test. Kimi is significantly better than Opus. That's the reason why I claimed it as Opus-5 level. Test done in @arena [video]
@synthwavedd Leo on x
I think Kimi K3 is going to shock some of the “the Chinese are 8 months behind the Western frontier” people
@kimmonismus @kimmonismus on x
Kimi k3 starts rolling out. Official release is imminent! Super freaking excited for the evals. Will it bear opus 4.8? Could be a real game changer, literally.
@thegenioo Hamza on x
I was wrong about Kimi K3 Just tried it in Claude Code and it is Fable 5 level for sure Far better than GPT-5.6 Sol at least at front-end It is really slow and consumes a lot of tokens but it is so good at completing its task, I like what I see here @Kimi_Moonshot [video]
@zackkorman Zack Korman on x
If Kimi 3 is actually better than Opus 4.8 then I just want to say it was good knowing you all. Stay strong through the cyber apocalypse.
@oluwaphilemon1 @oluwaphilemon1 on x
Kimi K3 is currently leading Claude Fable 5, GPT-5.6, and Grok 4.5 creating game demo. A mystery Arena .ai model called “Kivine” is reportedly Kimi K3. In this AI Game Watch, I break down the first game and 3D demos shared by creators on X. [video]
@theo @theo on x
Normally I don't comment on rumors, but if Kimi K3 actually beats out Opus 4.8 that's nuts Also hearing Opus 5 might drop?
@shanumathew93 Shanu Mathew on x
Reminder that Opus 4.8 came out at thee end of May. So frontier leadership went from 6+ months to 1-1.5 months. Unclear how much of this is because of wide distillation and how far internal models are at American labs but the direction is clear that China is catching up...
@theahmadosman Ahmad on x
Kimi K3 - 2.8T Parameters - 1M Context Length Benchmarks > GDPval-AA v2: 3rd place, ranks > directly below GPT 5.6 Sol Max & Fable 5 Max > AA-Briefcase: 2nd place, beats GPT 5.6 Sol Max, right below Fable 5 Max > BrowseComp: 1st place, beating GPT 5.6 Sol Max & Fable 5 Max [image…
@laschuk @laschuk on x
I put Kimi K3 up against Fable 5 in my benchmark. The test: clone Apple's homepage. Single prompt, one shot, zero help. Kimi K3: $0.44 (at left) Fable 5: $0.94 (at right) Same test. Same prompt. Half the price. Fable runs $10/$50 per MTok, the most expensive tokens on the [video]
@thegenioo Hamza on x
Kimi K3 is out. But it's not what I expected, not better than Fable or Opus 5 level. It is really good, persistent and determined, but tokens hungry and also very very slow. The designs are not out of this world on my test, but very clean, accurate and one shot. [video]
@jeffwang Jeff Wang on x
Chinese open source is no longer “6 months behind”, but it's also no longer “10% of the cost” either Congrats to Moonshot on Kimi K3, 2.8T, $3/$15 - same as Sonnet but initial feedback is great, we'll be sure to run FrontierCode on our side
@eliebakouch Elie on x
Kimi K3 (2.8T total parameter) will be the biggest open weight model ever and by far grok 4.5 is 1.5T just for comparison, deepseek v4 is 1.6T, opus and gpt 5.5 are said to be around 1.5T
@atomtanstudio @atomtanstudio on x
For everyone getting excited about Kimi K3. They are not cheap. [image]
@norwakar Diwakar Ray Yadav on x
Kimi K3 just went global and beta testers are saying it matches Opus-level output at nearly half the price. 🚀 China clearly isn't slowing down in this AI race. Overhyped or the real deal? Drop your take 👇 [image]
@kimmonismus @kimmonismus on x
Official In benchmarks, Kimi k3 is reportedly just behind GPT-5.6 and Fable 5, but ahead of Opus 4.8. However, in terms of price, it's on par with Sonnet 5. An absolute game changer. Insane! [image]
@chetaslua @chetaslua on x
Kimi K3 vs GPT-5.6 Sol Difference in taste is so stark , like if i swap kimi k 3 name with fable 5 people will trust it but this is not only about visuals , its also about function , you can see both achieve same result but the way to achieve is different kimi is more [video]
@nrehiew_ @nrehiew_ on x
Here are a few benchmark scores of K3 that have been officially confirmed This is a Fable/Sol class model that is strictly better than Opus 4.8 across the board at Sonnet pricing. Insane [image]
@theahmadosman Ahmad on x
Kimi K3: The Opensource AI Frontier [image]
@mweinbach Max Weinbach on x
At $3/$15 in out tokens eh, like i get why it's priced what it is and the compute needed but i have no desire to pay that given it doesn't seem token efficient
@mweinbach Max Weinbach on x
Kimi K3 being 2.8T parameter and 1M context is cool but show me the sparsity, show me the price How quick can the inference providers scale this to 200 tok/s This is what I care about!!!! Efficient HUGE models
@bridgemindai @bridgemindai on x
Kimi K3 just beat Fable 5 on the BridgeBench Horror House game test. I did not expect this at all. Kimi K3 is better than Fable 5 at game development and UI design. The two things Fable 5 was supposed to own. The 5x price increase suddenly makes sense. [video]
@sungkim Sung Kim on bluesky
🔹 Built for long-horizon agentic coding and self-evolving workflows — Tech blog: kimi.com/blog/kimi-k3 [image]
@wongmjane Jane Manchun Wong on x
Unfortunately, Kimi K3 is not even willing to discuss 9/11 [image]
@deredleritt3r Prinz on x
The most interesting question about Kimi K3 is whether it poses cyber risk. Kimi K3 benchmarks do not include a CyberGym score. Waiting for @AISecurityInst to bench this model.
@yzhang_cs Yu Zhang on x
K3 has now crossed the 1M context-length barrier, and DeepSeek's sparse attn has done the same. But what architecture will take us to 5M, 10M, or even longer? I'd always argue that fixed-state linear attn, especially GDN/KDA, is highly competitive here. Hybrid designs are
@yzhang_cs Yu Zhang on x
funny that K3 is great at making 3Blue1Brown videos, and the Quantile Balancing example in the blog took just a few shots to produce. https://kimi.com/...
@theojaffee Theo Jaffee on x
Registering my prediction of no widespread societal chaos after the open-sourcing of Kimi K3
@tszzl Roon on x
the world vision of open weights models running themselves, self replicating, training new versions of themselves (at least the kind of behavioral modifications that won't require massive compute scale), is really not very far away
@ryangreenblatt Ryan Greenblatt on x
Kimi K3 was significantly but not massively above my expectations. I'd tentatively guess it's similar in overall usefulness/usability to Opus 4.8 and in overall capability somewhat above Opus 4.8 (while also being somewhat more benchmaxxed). As a pretrain, it's probably somewhere
@yzhang_cs Yu Zhang on x
In the early morning, @nathancgy4, @Xinyu2ML, @Yulun_Du and I were preparing some showcases for the blog while watching the World Cup. The moment Argentina beat England, I felt something. I looked up, and saw the most unforgettable Beijing sunrise. So I took the photo. I knew [im…

Chronicles