Kimi K3 goes open with 2.8T parameters

Moonshot AI has unveiled Kimi K3, a 2.8 trillion-parameter model the company describes as the world’s first open 3T-class model. It ships with native vision, a 1 million-token context window, and is available now through Kimi.com, Kimi Work, Kimi Code, and the Kimi API. The company says full model weights will be released by July 27, 2026.

Kimi positions K3 as its most capable model yet, aimed at long-horizon coding, knowledge work, and reasoning. Moonshot says the model still trails the strongest proprietary systems, specifically Claude Fable 5 and GPT 5.6 Sol, but outperformed the other models it tested across its internal evaluation suite.

Technically, Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), alongside a Mixture of Experts design that activates 16 of 896 experts. Moonshot says these changes deliver about a 2.5x improvement in scaling efficiency versus Kimi K2.

In coding, the company highlights long engineering sessions, repository-scale work, terminal tool use, and tasks that mix software development with visual reasoning. It claims K3 built MiniTriton, a compact Triton-like GPU compiler, and in a separate 48-hour autonomous run designed and verified a chip on the Nangate 45nm library. According to Moonshot, that chip fits within 4 mm², closes timing at 100 MHz, and sustains more than 8,700 tokens/s decode throughput in simulation.

For research workflows, Moonshot says K3 reproduced I–Love–Q universal relations in computational astrophysics by reviewing 20+ papers, evaluating 300+ equations of state, generating 3,000+ lines of Python code, and producing an interactive HTML dashboard.

Recommended reading

Huawei’s Atlas 950 Packs 256 TB for AI

Moonshot is also adding Widgets and Dashboard to Kimi Work, two features meant to make K3 outputs more visual and persistent. Widgets can generate interactive components inside a chat and connect to local data or external plugins, while Dashboard collects those components into a persistent project view.

Kimi K3 pricing and availability

At launch, max thinking effort is the default mode, with low- and high-effort options planned for later updates. Moonshot says it is working with inference partners and open-source maintainers ahead of a broader rollout, and that more architectural and training details will arrive with a technical report.

API pricing is set at:

• $0.30/MTok for cache-hit input
• $3.00/MTok for cache-miss input
• $15.00/MTok for output

Moonshot says the official API, powered by Mooncake’s disaggregated inference architecture, achieves a cache hit rate above 90% in coding workloads.