Moonshot AI has released Kimi K3, a 2.8 trillion parameter open-weight model now ranked as the largest open model ever built. Independent evaluations place its intelligence on par with Anthropic’s Opus 4.8 and GPT-5.5, while its frontend coding abilities took the #1 spot on Arena.AI’s human preference leaderboard. The model is live today on the Kimi platform and API, with full model weights scheduled for open release on July 27, 2026.
Why does Kimi K3 matter for open AI?
For most of the last three years, open models have trailed the best closed-source systems from OpenAI and Anthropic by a noticeable margin. A 2.8 trillion parameter open-weight model that matches or surpasses frontier systems on several real-world benchmarks changes that dynamic. Moonshot’s move follows a pattern among leading Chinese AI labs, which have increasingly open-sourced their most capable models. As Reuters noted, this strategy allows companies to "showcase their technological capabilities and expand developer communities as well as their global influence, a strategy likely to help China counter U.S. efforts to limit Beijing’s tech progress."
The model lands at a moment when enterprise technology leaders are questioning the cost and data privacy implications of paying for closed-source AI. The availability of a frontier-class open model is likely to accelerate a trend already underway: businesses pulling capable open alternatives inward, fine-tuning them on proprietary data, and avoiding the lock-in that comes with API-only access to the most powerful models.
What’s new and how does Kimi K3 work?
Kimi K3 arrives with a spec sheet that makes it a dramatic outlier. The 2.8 trillion total parameters represent roughly 75% more capacity than DeepSeek’s V4 Pro, which the company’s own timeline chart shows at roughly 1.6 trillion parameters. The model also features a 1-million-token context window and native multimodal input that accepts text and images.
Moonshot achieved this scale through two architectural innovations that it published openly ahead of the launch. Kimi Delta Attention (KDA) is a hybrid linear attention mechanism that the company says enables up to 6.3x faster decoding in million-token contexts. Attention Residuals (AttnRes) act as a drop-in replacement for standard residual connections, delivering roughly 25% higher training efficiency at only 2% additional cost. Community readers of the technical blog have also identified the use of LatentMoE with 16 activated experts out of 896, an activation ratio under 2%, a new activation function called SiTU (Sigmoid Tanh Unit), and quantile load balancing. These choices stack to create a model that scales more than 2x over prior Kimi architectures.
On the practical side, the model is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens dropping to $0.30 per million. A blended estimate at 80% input and 20% output puts the effective cost around $5.40 per million tokens, compared to $9 for Opus 4.8 and $10 for GPT-5.5. Moonshot also contributed a KDA prefix caching implementation directly to vLLM, ensuring day-0 runtime support for the open release. The company itself recommends deployment on supernode configurations with 64 or more accelerators for best inference efficiency.
What do the benchmark numbers show?
- 2.8 trillion total parameters, makes Kimi K3 the largest open-weight model ever released, roughly 75% larger than DeepSeek V4 Pro.
- 57 on the Artificial Analysis Intelligence Index, comparable to Opus 4.8 and GPT-5.5, behind Fable 5 and GPT-5.6 Sol (source).
- 1,687 on GDPval-AA v2, third overall, ahead of Opus 4.8 (1,600) on a benchmark covering 44 occupations and 9 industries (source).
- #1 on Arena.AI Frontend Code Arena with 1,679 points, a 76% pairwise win rate, versus 63% for Fable 5 and 58% for GPT-5.6 Sol.
- #1 on AutomationBench-AA with 53%, an independent agentic benchmark from Artificial Analysis.
- 91.2 on BrowseComp, a state-of-the-art score on a long-horizon information-seeking benchmark.
- Open weights promised by July 27, 2026, full model weights to be released under an open license.
"parameters are like neural connections in the human brain, and nearly 3 trillion of them means the model can store more knowledge and patterns in its brain, understand more, think deeper, and answer more accurately."
This explanation came from a Moonshot AI executive speaking to Xinhua, the state news agency, which framed K3 as a national milestone. Liu Tieyan, dean of the Zhongguancun Academy in Beijing, added that a wave of Chinese open-source models has moved from isolated breakthroughs to collective advancement.
What comes next for Kimi K3?
The July 27 weight release is the immediate milestone. Once the full model is downloadable, any organization with sufficient GPU infrastructure will be able to fine-tune, distill, or self-host a frontier-class model without an API contract. Moonshot is also reportedly raising a fresh funding round at a $31.5 billion valuation, following a $2 billion raise at a $20 billion valuation in May, signaling that investors expect Kimi K3 to fuel an ecosystem rather than a one-time headline.
The company is already demonstrating what agentic capabilities look like at this scale. In a 48-hour autonomous chip design demo, K3 completed the full pipeline for a 4-square-millimeter functional chip design that achieved timing convergence at 100 MHz and could decode more than 8,700 tokens per second in simulation. In computational astrophysics, K3 reportedly reproduced the universal I-Love-Q relation, a task that typically takes a senior researcher one to two weeks, in approximately two hours, reading and cross-validating more than 20 papers along the way. These demos hint at the long-horizon autonomous coding and research workflows Moonshot is betting on as the next competitive frontier.
What does this mean for your business or site?
For business owners, marketers, and anyone who builds on top of AI, the arrival of a truly frontier-class open model changes the math. You are no longer choosing between a handful of API gateways. You now have an option that you can self-host, fine-tune on your own data, and integrate without per-token prices that scale with usage. The current pricing is already competitive with mid-tier closed models while delivering Opus 4.8-class performance, and the open weights will let you run it entirely on your own terms after July 27.
For anyone who produces content, runs a site, or cares about how AI is changing search, Kimi K3 is also a reminder that the model landscape is shifting fast. The models that power AI overviews, chatbots, and agentic search are no longer the exclusive domain of two Silicon Valley labs. We explored the durable SEO strategies that hold up as AI reshapes discovery in The Long-Term Game: SEO Strategies That Hold Up in AI Search. And when we covered Google’s own guidance for the AI overview era, the core message was clear: generic, interchangeable content is the real liability. Read more in Google’s Generative AI Search Guide Reads Like a Warning About Generic Content.
Pay attention to open model releases like this one. They are not just benchmarks and parameter counts. They are the raw material that will power the next wave of tools your customers are already using.
What is the bigger picture for open vs closed AI?
Kimi K3 is not just a larger model. It is a signal that the gap between open and closed frontier intelligence has narrowed to a point where businesses must factor it into their AI strategy. When a 2.8 trillion parameter model that rivals Opus 4.8 can be downloaded in a matter of weeks, the question is no longer whether open models will catch up. The question is what you will build when they do.
A 2.8 trillion parameter open model that rivals closed frontier systems is now real, and its weights will be free in two weeks.
FAQ
What is Kimi K3?
Kimi K3 is a large language model developed by Beijing-based Moonshot AI. It contains 2.8 trillion parameters, making it the largest open-weight AI model ever released, with a 1-million-token context window and native multimodal input for text and images. The model is available now through the Kimi chatbot and API, with full model weights scheduled for open release on July 27, 2026.
How does Kimi K3 compare to Opus 4.8 and GPT-5.5?
Independent evaluations from Artificial Analysis place Kimi K3 at an intelligence score of 57, comparable to Opus 4.8 and GPT-5.5. On GDPval-AA v2, Kimi K3 scored 1,687, ahead of Opus 4.8 at 1,600. It also reached #1 on Arena.AI’s Frontend Code Arena with a 76% pairwise win rate.
How much does Kimi K3 cost to use?
The Kimi K3 API is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens at $0.30 per million. A blended estimate assuming 80% input and 20% output puts the effective price around $5.40 per million tokens, compared to $9 for Opus 4.8 and $10 for GPT-5.5.
Related coverage
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.