Moonshot AI releases Kimi K3 — open-weight frontier at 2.8T parameters
In mid-July 2026, Moonshot AI unveiled Kimi K3 — per the vendor, a model with 2.8 trillion parameters (2.8T), announced in the middle of a livestream. The open weights followed on July 27, 2026. Moonshot positions K3 as an open frontier model with strong agentic capabilities. The parameter count is a vendor figure and, so far, is not confirmed by independent benchmarks.
What is known so far
- Model: Kimi K3 from Moonshot AI (China), successor to the K2 line.
- Size: 2.8 trillion parameters — a vendor figure, likely a mixture-of-experts with a much smaller active parameter count.
- Launch: mid-July 2026, announced live on stream.
- Open weights: freely downloadable from July 27, 2026.
- Positioning: open frontier model focused on agentic tasks.
Where things stood
Moonshot AI had already established itself as a serious open-weight vendor with the Kimi K2 line — most recently with K2 Thinking and the point releases K2.5 and K2.6. The ambition there was already to ship open models at frontier level, with a large context window and solid coding performance. Together with DeepSeek and Qwen (Alibaba), Moonshot formed the spearhead of the Chinese open-weight labs that compete with Western vendors mainly on price and openness.
The difference from the large Western labs remained the availability of the weights: anyone who wanted to self-host a model, fine-tune it, or place it in a closed environment depended on exactly these open models — the top models from Anthropic, OpenAI, and Google are accessible only via API.
What now applies
1. An open-weight model moves into frontier territory by the numbers. At a reported size of 2.8T parameters, Kimi K3 nominally sits among the largest known models. Important context: for mixture-of-experts architectures, the raw parameter count says little about the actual compute per request, because only a fraction of the parameters is active per token. The number is a marketing-relevant headline figure, not a performance proof.
2. The weights are open. The point that matters in practice is not the size but that K3 is freely downloadable from July 27. That allows self-hosting, integration into your own pipelines, and operation without API lock-in — provided you have the required hardware, which is substantial for a model of this scale.
3. Agentic focus. Moonshot explicitly markets K3 for agentic tasks — multi-step workflows with tool use, as common in coding assistants and autonomous workflows. Whether real-world performance lives up to the announcement will have to be shown by independent tests.
Reading
The actual news value lies not in the parameter count but in the pattern. Within a few weeks, Kimi K3, DeepSeek V4-Pro, Qwen 3.7-Max, and MiniMax M3 have all released open models close to frontier level. That raises pressure on the closed labs in two places: on price — open models can be self-hosted or served by cheap providers — and on openness, which is increasingly an argument for regulated or data-sensitive deployments.
Caution about the figures remains warranted. “2.8 trillion parameters” is a vendor figure with no disclosed architecture details. Without independent benchmarks and without the active parameters per token, the number is mainly a headline. A reliable assessment needs community evals, which will only emerge after the weight release on July 27.
What you can do now
If you are evaluating self-hosting or open-weight strategies: Kimi K3 belongs on your watch list from the weight release — but wait for independent benchmarks before turning vendor claims into decisions. Details on the family are in our glossary entry on Kimi K3.
If you select models in a client context: treat the open Chinese frontier models as a real option for price and data-sovereignty cases — but check license terms, hosting costs, and compliance requirements concretely instead of following the parameter count.