Quick Facts
- Moonshot AI released Kimi K3’s full weights on Hugging Face on July 27, 2026, under a custom license that requires separate commercial agreements for Model-as-a-Service deployments above certain revenue thresholds.
- Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window, scoring 57.1 on Artificial Analysis — fourth overall, above Claude Opus 4.8 and Grok 4.5.
- API pricing starts at $3 per million prompt tokens and $15 per million output tokens, roughly three to four times the cost of its predecessor, Kimi K2.6.
Moonshot AI has published the full weights for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model the company calls the world’s first open 3T-class model. The weights went live on Hugging Face on July 27, 2026, roughly 11 days after Moonshot unveiled the model at the World AI Conference in Shanghai.
For enterprises, the release carries a significant caveat: the model ships under a custom Kimi K3 License, not a standard open-source license. Organizations running a Model-as-a-Service product above specified revenue thresholds must negotiate a separate commercial agreement with Moonshot. The largest deployments must also display visible attribution to Kimi K3 inside their products.
Where It Performs and Where It Doesn’t
The benchmark numbers are strong. Kimi K3 outperforms Claude Fable 5 on Terminal-Bench (88.3 vs. 84.6) and SWE Marathon (42.0 vs. 35.0). On BrowseComp, which tests real-world research and browsing tasks, K3 scores 91.2, above Claude Fable 5 at 88.0 and GPT-5.6 Sol at 90.4.
The model debuted at 57.1 on Artificial Analysis, placing it fourth overall and ahead of Claude Opus 4.8 (55.7), Grok 4.5 (53.8), and GLM-5.2 (51.1). On per-task cost, K3 runs about $0.94 per weighted Intelligence Index task, compared with $1.04 for GPT-5.6 Sol and $2.75 for Fable 5 with fallback.
One problem stands out for production use. On the AA-Omniscience hallucination benchmark, K3’s hallucination rate climbed from 39% to 51% even as its accuracy improved from 33% to 46%. For legal, financial, medical, or factual research workflows where an incorrect answer is worse than no answer, that rate is a material risk.
The model is also verbose. Kimi K3 generated 130 million output tokens across the Artificial Analysis evaluation suite, against a 63 million average for other models. Real-world inference costs will depend heavily on how much reasoning-token burn each workload triggers.
Background: A Comeback Play
Moonshot AI was founded in 2023 by Yang Zhilin, a Tsinghua University graduate who previously did research at Google and Meta. The company raised roughly $1.5 billion across multiple funding rounds, with its valuation rising from $2.5 billion to $4.3 billion by early 2026.
The company’s fortunes shifted after DeepSeek released its low-cost R1 model in January 2025. Kimi, which had ranked third in monthly active users in China, fell to seventh. Moonshot responded by pivoting toward open-weight releases, starting with Kimi K2 in July 2025 and continuing with K2.5 in January 2026. Kimi K3 is the latest step in that strategy.
What Enterprises Should Do First
The distinction between open weights and open source is sharpening. Downloading and modifying Kimi K3 is technically feasible. Determining the legal conditions for commercial deployment requires reading the license carefully before any infrastructure investment.
Organizations planning internal deployments — supporting developers, researchers, legal teams, or productivity workflows — face substantially fewer restrictions under the published terms. Those planning to build externally facing AI products or services on top of K3 need to review whether their revenue profile triggers the commercial agreement requirement.
Moonshot published not only the weights but also supporting infrastructure, including attention kernels, MoE communication libraries, and agent tooling. That breadth of release earned praise from developers. However, running a 2.8-trillion-parameter model at production scale demands significant hardware investment that narrows the realistic self-hosting pool to well-resourced organizations.
Read more: Kimi K3’s full weights are here, but they’re ‘open’ with a caveat: What enterprises should know
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.