Kimi K3 promises a 2.8-trillion-parameter open-weight release

Moonshot AI has released Kimi K3, a 2.8-trillion-parameter model aimed at long-running coding, research and knowledge-work agents. The Chinese AI company describes it as the first open model in the three-trillion-parameter class, while acknowledging that its overall performance still trails Claude Fable 5 and GPT 5.6 Sol.

That scale makes the release unusual, but the launch needs one qualification. K3 is available now through Kimi’s consumer products, coding agent and API, yet Moonshot says the full weights will arrive by 27 July. On launch day there was no K3 repository, model card or confirmed licence on Moonshot’s GitHub and Hugging Face organisations. Developers can test the hosted model but cannot independently inspect or deploy the promised open-weight release. Reuters therefore has grounds to report Moonshot’s claim that K3 is the world’s largest open AI model, although the claim will only become fully testable once the company publishes the weights, licence and technical report.

K3 uses a mixture-of-experts architecture, so it does not engage all 2.8 trillion parameters for every token. Moonshot says its Stable LatentMoE design selects 16 of 896 experts during inference. The model combines this sparse routing with Kimi Delta Attention and Attention Residuals, two mechanisms intended to improve information flow across long sequences and deep networks. Moonshot claims the resulting design delivers roughly 2.5 times the scaling efficiency of Kimi K2.

The practical constraint is infrastructure. Moonshot recommends supernode deployments with at least 64 accelerators, even after using quantisation-aware training with MXFP4 weights and MXFP8 activations. K3 may be open enough for cloud providers, research laboratories and large enterprises to operate, but its size puts local deployment well beyond ordinary developer hardware. Open weights will improve control and competition without making this a desktop model.

The hosted version is easier to assess. K3 exposes a one-million-token context window through an OpenAI-compatible API and supports text, images, video, structured output, tool calling and dynamic tool loading. It always runs in thinking mode at launch, with only the maximum reasoning level currently available. Moonshot charges $0.30 per million cached input tokens, $3 per million uncached input tokens and $15 per million output tokens. The company says its disaggregated Mooncake inference system achieves cache-hit rates above 90% on coding workloads, a claim that would make the low cached-input rate commercially significant for large, stable repositories.

Moonshot has positioned K3 around tasks that run for hours rather than short chat responses. Its launch material includes GPU-kernel optimisation, a small Triton-like compiler, chip design, scientific computing and research reports assembled through thousands of searches and terminal operations. One demonstration describes a 48-hour run in which K3 designed and verified a small accelerator using open-source electronic design tools. These examples are provider-selected case studies rather than independent evaluations, but they show the intended product: an agent that can retain a long plan, work across tools and keep revising its output.

The benchmark presentation deserves similar care. Moonshot reports competitive coding and agentic results across DeepSWE, Terminal-Bench, Program Bench, FrontierSWE, OfficeQA Pro, MCP Atlas and BrowseComp. K3 records 67.3 on DeepSWE with the mini-SWE-agent harness and 90.4 on BrowseComp when given the full one-million-token context without compaction. Comparisons across the wider table use several different agent harnesses, including Kimi Code, Claude Code and Codex, and some scores come from third parties while others come from Moonshot’s own runs. They are useful evidence for selecting models to test, rather than a clean ranking of model intelligence.

Moonshot’s own limitations section is more informative than many launch claims. K3 expects an agent harness to preserve its complete reasoning history between turns. Dropping that history, or switching to K3 partway through a session started with another model, can make generation unstable. The company also warns that training for difficult, long-running tasks can make the model overly proactive when instructions are ambiguous. Teams giving it terminal, browser or business-system access will need explicit boundaries, narrow permissions and reliable checkpoints.

Those caveats define the engineering opportunity. K3 gives infrastructure providers a very large open-weight model to serve and optimise, while the API gives application teams immediate access without building a 64-accelerator cluster. Its long context, tool controls and cached-input pricing suit repository-scale coding and research agents, provided workloads can tolerate a model that always reasons at maximum effort and may require tighter supervision than the best proprietary systems.

Kimi K3 also adds pressure to the frontier-model market from a different direction. Moonshot is not claiming an outright win over the strongest closed models. It is offering near-frontier capability, an unusually large architecture, a hosted price that remains competitive and a commitment to release the weights shortly after launch. The important evidence will arrive after 27 July, when independent teams can examine the licence, reproduce benchmark runs and discover what it costs to serve a 2.8-trillion-parameter model outside Moonshot’s own infrastructure.

Sources: Reuters, Moonshot AI Kimi K3 technical blog, Kimi K3 API guide, Kimi K3 pricing.