Kimi K3: Benchmarks, Price and Open Weights

On 16 and 17 July 2026 the Chinese provider Moonshot AI released its new model Kimi K3, usable via API and on kimi.com. This is not a minor announcement: according to VentureBeat, Kimi K3 is the largest open-source model ever released and rivals the strongest US systems. For anyone using AI in day-to-day marketing and work, it pays to look soberly at the numbers before the headlines frame them.

I am treating this as a release analysis: what is inside the model, how the benchmark comparison against Fable 5 and GPT-5.6 Sol turns out, why open weights matter strategically, and where I honestly still wait and see. Which models we actually use in the agency we keep transparent on our AI in numbers page.

What Kimi K3 is

The technical specs are unusually large. According to Tom's Hardware, Kimi K3 has around 2.8 trillion parameters, making it the largest open-weight model announced so far. The context window spans one million tokens (exactly 1,048,576), and according to Moonshot the model understands image and video natively.

Kimi K3 has been available since the release weekend via API and on kimi.com. The full weights are, according to Moonshot, announced for 27 July 2026, with a technical report to follow. Notable is the geopolitical framing that Tom's Hardware stresses: with models like this, China is visibly working around US compute restrictions. That the largest open model comes of all places from China is therefore more than a side note. CNBC also frames the release as a direct challenge to OpenAI and Anthropic.

Important for context: "largest open-source model" is first of all a claim by Moonshot and the reporting outlets, not an independently verified award. The parameter count alone also says little about actual quality. That is what benchmarks are for.

The benchmarks: 2nd to 4th in the frontier class

The central yardstick is the Intelligence Index from Artificial Analysis. There Kimi K3 scores 57 and lands 4th of 189 listed models, behind Fable 5 (60) and GPT-5.6 Sol (59). Moonshot itself states in internal evaluations that the model sits "behind only Fable 5 and GPT-5.6 Sol". That is a vendor claim, but it roughly matches the order at Artificial Analysis.

It gets more interesting in the more application-oriented tests. In GDPval-AA v2, which according to Artificial Analysis maps real work across 44 professions and 9 industries, Kimi K3 reaches 1,687 points for 3rd place, behind Fable 5 Max (1,815) and Sol Max (1,747.8) but ahead of Claude Opus 4.8 Max (1,600). An open-weight model that overtakes established closed systems on simulated professional work is remarkable.

The picture is even clearer in AA-Briefcase, a private Artificial Analysis benchmark for long, agentic task chains. Here Kimi K3 scores 1,527 for 2nd place. That beats Sol Max (1,495) and trails only Fable 5 Max (1,587). According to Artificial Analysis this corresponds to an Elo jump of 732 points over the predecessor Kimi K2.6, an unusually large step between two generations.

On top of that come two single scores that stand out. On BrowseComp, the outlet officechai reports 91.2, a state-of-the-art value. And in the Frontend Code Arena, Kimi K3 even beats Fable 5, according to a Tom's Hardware headline. Both are individual best marks, not across-the-board leadership, but they show the model plays right at the front in specific disciplines.

Price-performance: the real lever

Pure intelligence numbers are one side. The other, often more important in practice, is price. And this is where Kimi K3 gets genuinely interesting.

According to the release figures, the model costs 3 dollars per million input tokens on a cache miss, only 0.30 dollars for cached tokens and 15 dollars per million output tokens. What matters is the calculation Artificial Analysis derives from this: converted to a weighted Intelligence Index task, Kimi K3 costs around 0.94 dollars. For comparison, Artificial Analysis lists 1.04 dollars for GPT-5.6 Sol Max and 2.75 dollars for Fable 5 with fallback. That gives Kimi K3, according to Artificial Analysis, the best price-performance ratio in the frontier class.

This is the point where the discussion turns for me. A model that ranks 2nd to 4th in the relevant leaderboards while charging a fraction of the quality leader's cost changes the math for many workflows. Anyone who followed the expensive comeback of Fable 5 and the benchmark and cheating debate around GPT-5.6 Sol can see it: the frontier class has become tighter and more nuanced than the headlines from individual providers suggest.

What open weights mean strategically

The second big point alongside price is the open weights, announced for 27 July 2026. The difference from closed models is both practical and strategic.

Open weights enable three things that are impossible with pure API models. First, self-hosting: the model can run on your own or rented infrastructure, without every request going through the provider's server. Second, fine-tuning: you can specialize the model on your own data and tasks. Third, independence: whoever holds the weights no longer depends on a provider keeping the model available, holding the price stable or not restricting access. How quickly a frontier model can disappear was just shown by the shutdown episode around Fable 5.

For companies that seriously build AI into their processes, this is a real argument. It shifts the question from "which provider do I book?" to "which model do I run myself?". This exact trade-off is part of what I discuss with clients in AI consulting, often combined with the question of how it all connects to clean AI automation in Google Ads.

My take

This is my personal view, clearly framed as an opinion and not as an external fact.

For me, Kimi K3 is currently the best price-performance model in the frontier class. Ranking 2nd to 4th in the relevant leaderboards at a fraction of the Fable cost is a combination that has rarely existed before. If the Artificial Analysis numbers hold up, it is a very attractive offer for many tasks.

I find the open weights at least as important strategically as the benchmarks. Self-hosting, fine-tuning and no vendor lock-in are arguments that go beyond raw model quality and matter especially for European companies.

Still, I deliberately wait and see on three points. First, practical quality beyond the benchmarks: at the time of this article the model is a few hours old, and numbers are no substitute for real work over weeks. Second, the weights release itself: 27 July is announced, but announced is not shipped. Third, European compliance: with a Chinese provider, data processing via the API raises questions you have to clarify beforehand. With self-hosting this point largely falls away, which makes the open weights even more valuable. My conclusion: great potential, but I judge it by practice, not by the headline.

Sources

• VentureBeat: China's Moonshot AI releases Kimi K3, the largest open-source model ever
• CNBC: Moonshot AI Kimi K3 model, OpenAI, Anthropic, China
• Tom's Hardware: Moonshot releases 2.8 trillion parameter Kimi K3
• Simon Willison: Kimi K3
• Artificial Analysis: model rankings and Intelligence Index
• BrowseComp value (91.2) and state-of-the-art note: officechai (outlet, no direct link in the dossier)

As of: 17 July 2026

Frequently asked questions

What is Kimi K3?

Kimi K3 is the new AI model from Chinese provider Moonshot AI, released on 16 and 17 July 2026. With 2.8 trillion parameters it is, according to VentureBeat and Tom's Hardware, the largest open-source model ever announced. It has a one-million-token context window and, according to Moonshot, understands image and video natively. It is usable via API and on kimi.com; the full weights are announced for 27 July 2026.

How does Kimi K3 perform in the benchmarks?

In the Artificial Analysis Intelligence Index Kimi K3 scores 57, ranking 4th of 189 models, behind Fable 5 (60) and GPT-5.6 Sol (59). In the work benchmark GDPval-AA v2 it reaches 1,687 for 3rd place, in the agentic AA-Briefcase it hits 1,527 for 2nd place and beats Sol Max (1,495) there. On BrowseComp, officechai reports 91.2 as state of the art.

What does Kimi K3 cost?

According to the release figures, Kimi K3 costs 3 dollars per million input tokens on a cache miss, 0.30 dollars for cached tokens and 15 dollars per million output tokens. Converted by Artificial Analysis to a weighted Intelligence Index task, Kimi K3 lands at around 0.94 dollars, versus 1.04 dollars for GPT-5.6 Sol Max and 2.75 dollars for Fable 5, which Artificial Analysis rates as the best price-performance ratio in the frontier class.

What do open weights mean for Kimi K3?

Moonshot has announced the full weights for 27 July 2026. Open weights allow self-hosting, fine-tuning on your own data and reduce dependence on a single provider. My honest caveat: until the weights actually ship, until practical quality beyond the benchmarks is proven and until European compliance questions around a Chinese provider are clarified, this remains an announcement, not a settled fact.