New model released: Kimi K3 from Moonshot AI

On July 17, 2026, Moonshot AI introduced Kimi K3, which the company describes as its most capable model to date and the first openly available model in the 3-trillion-parameter class. Kimi K3 has 2.8 trillion parameters, a one-million-token context window, and native vision capabilities.

The key facts

• Size and architecture: Kimi K3 is built on two new architectural components called Kimi Delta Attention and Attention Residuals, and uses a refined Mixture-of-Experts structure that activates 16 of 896 experts at a time. Compared to its predecessor Kimi K2, Moonshot says scaling efficiency improves by roughly 2.5x.
• Performance compared: According to Moonshot, Kimi K3's overall performance still trails the proprietary frontier models Claude Fable 5 and GPT 5.6 Sol, but it consistently outperforms every other tested model on the company's own evaluation suite. Independent coverage, including from Bloomberg and CNBC, frames Kimi K3 as a significant narrowing of the gap between Chinese models and leading US systems.
• Pricing: Through the Kimi API, usage costs $0.30 per million input tokens on a cache hit, $3.00 on a cache miss, and $15.00 per million output tokens.
• Availability: Kimi K3 is available today via Kimi.com, the Kimi Work desktop app, Kimi Code, and the Kimi API. Full model weights were made publicly available on July 26, 2026, a day earlier than originally announced.

Benchmark comparison

At launch, Moonshot published a comparison chart positioning Kimi K3 against the leading proprietary models.

Coding and agentic capabilities

Moonshot specifically highlights long-horizon, agentic coding tasks: Kimi K3 is said to independently navigate large repositories, orchestrate terminal tools, and sustain long working sessions with minimal human oversight. In one documented case, the model completed a scientific coding task reproducing universal relations in astrophysics in about two hours, a task Moonshot says would typically take an experienced researcher one to two weeks. As a standalone project, Kimi K3 also built MiniTriton, a compact GPU compiler that matched or outperformed the established Triton framework on several benchmarks on an NVIDIA L20 GPU.

Knowledge work

On complex knowledge-work tasks too, Moonshot's internal evaluations show consistent gains over previous models, for example in producing interactive research websites, industry analyses, and scientific data workflows.

A brief note on the political context

The release came amid heightened tension between the US and China over AI. According to reports, the White House Office of Science and Technology Policy alleged that Moonshot was running an internal distillation operation against Claude Fable 5 using restricted Nvidia chips. Moonshot has not publicly commented on the allegation; critics of the claim point out that K3 testing reportedly began before Fable 5 was even released.

Limitations, according to Moonshot

Moonshot lists three limitations in its own announcement. First, Kimi K3 is sensitive to whether its full thinking history is correctly passed back to the model, so switching to K3 mid-session from a different model can lead to unstable results. Second, when facing ambiguous user intent, the model tends toward a high degree of initiative and can make unexpected decisions on the user's behalf, which is why Moonshot recommends setting more explicit behavioral constraints in the system prompt where needed. Third, Moonshot acknowledges a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.