Moonshot AI has released Kimi K3, and the headline number is almost comically large: 2.8 trillion parameters.
That makes K3 the first model Moonshot describes as an open 3T-class system. It has native vision, a one-million-token context window, and a sparse mixture-of-experts architecture that activates 16 of 896 experts. It is available through Kimi's consumer, work, coding, and API products now, while Moonshot says the full model weights will arrive by July 27.
The size will dominate the first round of coverage. It should not dominate the serious discussion.
Kimi K3 matters because it is another attempt to break the false choice that has settled over frontier AI: either use the best closed model through somebody else's API, or run a smaller open model that gives you control but asks you to accept a capability gap. Moonshot is arguing that the gap is becoming small enough to make that tradeoff uncomfortable for the incumbent labs.
That argument is plausible. It is not proven yet.
What Moonshot Actually Released
Kimi K3 is a natively multimodal reasoning model built for coding, knowledge work, visual creation, and long-running agent tasks. Moonshot says its architecture combines Kimi Delta Attention and Attention Residuals with a Stable LatentMoE design.
The jargon has a practical point. Kimi Delta Attention is intended to make attention scale more efficiently over long sequences. Attention Residuals let the model retrieve representations across depth instead of simply accumulating every layer in the same way. The sparse expert design means only a small fraction of the total parameter count is active for a given token.
That is how a 2.8T-parameter model can be technically meaningful without requiring every parameter to participate in every step. Moonshot claims the architecture and training recipe deliver roughly 2.5 times the scaling efficiency of Kimi K2. It also says K3 was trained with MXFP4 weights and MXFP8 activations in mind, an important signal that the company is thinking about inference hardware rather than publishing a number that nobody can serve.
For users, the simpler list matters more:
- one million tokens of context;
- text, image, and video understanding;
- long-horizon coding and terminal-tool use;
- multi-agent research and knowledge-work workflows;
- API access through the
kimi-k3model; - full weights promised by July 27, 2026.
At launch, the official API costs $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens. That is not bargain-bin inference, particularly on the output side. It is frontier-model pricing with an aggressive cache story.
The Benchmarks Are Strong—and Still Vendor Benchmarks
Moonshot's published table puts Kimi K3 close to the strongest proprietary models across coding, tool use, browsing, office work, reasoning, and vision. Some results are genuinely eye-catching.
K3 reports 88.3 on Terminal-Bench 2.1, just behind GPT-5.6 Sol at 88.8 in Moonshot's table. It reports 81.2 on FrontierSWE, ahead of the 71.3 listed for Sol, and 84.2 on MCP Atlas, close to the 84.7 reported for Claude Fable 5. It also posts strong results on BrowseComp, Automation Bench, SpreadsheetBench 2, and multimodal document tasks.
But model-launch tables are not neutral scoreboards. Harness choice matters. Reasoning effort matters. Tool configuration matters. Context management matters. Moonshot's own footnotes show that different models were tested through different coding harnesses, with some competitor scores pulled from external leaderboards and others run by Moonshot.
That does not make the results useless. It makes them the start of evaluation, not the end.
Moonshot is unusually candid in a few places. It says K3 still trails Claude Fable 5 and GPT-5.6 Sol overall. It warns that K3 was trained to preserve its thinking history and can become unstable if an agent harness does not pass that history back correctly, or if a session switches to K3 halfway through. It also acknowledges a noticeable user-experience gap against the leading proprietary models.
Those caveats make the release more credible. They also tell platform teams exactly what to test: not just whether K3 solves a benchmark, but whether it survives a long session, a context compaction, a tool error, a model fallback, and a mid-task handoff.
“Open” Is a Promise Until the Weights Ship
The most important sentence in the announcement is not the parameter count. It is the date attached to the weights.
Kimi K3 can be used today, but the full weights are promised by July 27. Until those files, the license, the model card, and enough serving detail are available for independent inspection, K3 is an API-accessible model with an open-weight release plan.
That distinction matters.
Open weights let researchers inspect behavior, infrastructure companies build their own serving stacks, enterprises choose where inference runs, and developers fine-tune or adapt a model without waiting for a provider roadmap. They also make it possible to test the vendor's claims outside the vendor's preferred harness.
The first real K3 milestone is the product launch. The second is the weight release. The third is whether the community can serve the model reliably enough for anyone beyond hyperscalers, specialist inference providers, and very well-funded research groups.
A 2.8T sparse model may be open, but it will not be casual to operate. Download size, quantization quality, expert routing, interconnect bandwidth, memory layout, and distributed inference all become part of the product. Open weights remove a permission boundary. They do not remove physics or the cloud bill.
That is why the ecosystem response matters more than the first-day demo reel. If inference providers support K3 well, open-source harnesses preserve its reasoning history correctly, and quantized deployments retain quality, Moonshot will have created a real alternative. If the model is technically open but operationally impractical, most users will still consume it through a hosted API.
K3's Best Work Is Already Agent Work
Moonshot is not presenting K3 as a better chatbot. The release is organized around work.
The company says K3 optimized GPU kernels, helped build a compact Triton-like compiler, iterated on games by reading live screenshots, produced research sites from thousands of searches and data pulls, coordinated more than 20 subagents for scientific analysis, and edited video from dozens of source clips.
These are vendor demonstrations, so they deserve the same caution as the benchmark table. Still, the shape of the examples is right. Useful frontier systems are moving away from one prompt and one answer. They need terminals, browsers, files, visual feedback, parallel workers, durable context, and a place to run.
That shift is bigger than Kimi K3. Every major model launch now describes some version of an operating loop: inspect, plan, call tools, observe, revise, and continue. Model labs are converging on agent infrastructure because that is where stronger reasoning turns into economically useful output.
The model is becoming one component in a much larger runtime.
What This Means for Clanker Cloud
Kimi K3 is exactly why Clanker Cloud is designed around model choice rather than model loyalty.
If K3 proves excellent at long-repository coding, research, or visual iteration, an agent workspace should be able to route those tasks to it. If another model is more reliable for a production migration or more conservative around security work, the workspace should route differently. If open weights make private deployment practical for a particular organization, that should be an option too.
But routing is only useful when the operating boundary is clear.
A million-token context window does not grant trustworthy cloud context. A strong coding score does not tell the model which AWS account is production. Native vision does not authorize a click. Open weights do not protect a kubeconfig. An agent still needs isolated space for exploratory work, local custody of credentials, explicit tools, audit evidence, and human review before a destructive change.
Clanker Cloud's sandboxes give agents a hosted workspace for preparation and testing. The local-first desktop app and open-source Clanker CLI provide live infrastructure context without handing the model provider control of the user's credentials. MCP exposes tools and state in a standard way. Review-before-apply keeps high-impact work behind an approval boundary.
K3 can make that loop smarter. It cannot replace the loop.
The Bottom Line
Kimi K3 looks like a serious frontier release. The combination of a 2.8T sparse architecture, native multimodality, a million-token context window, strong vendor-reported agent benchmarks, and a scheduled open-weight release is enough to make every closed-model provider pay attention.
It is also too early for victory laps.
The weights have not shipped yet. The technical report is still coming. Independent evaluators have had little time with the model. Long-running reliability, harness compatibility, serving economics, and real-world safety will matter far more than the first leaderboard screenshots.
Still, this is the healthy kind of pressure. Kimi K3 forces the frontier market to answer a useful question: if an open-weight model can get close enough on capability, how much control, portability, and provider independence are teams willing to trade away for the final few points of polish?
That is a much better argument for the industry to have than another week of treating one closed API as destiny.
Sources
Give your agent live infrastructure context
Download Clanker Cloud, expose the local MCP surface, and let coding agents work from current cloud, Kubernetes, GitHub, and cost state instead of guesses.