Kimi K3 is a 2.8-trillion-parameter bet on agents that can keep working

The easiest way to read Kimi K3 is as another very large model. Moonshot says it has 2.8 trillion parameters, a number so large that it dominates almost every first impression of the launch.

The more useful reading is about endurance. Kimi K3 is built to keep context, use tools, inspect its own output, and continue working across coding and knowledge tasks that take hours rather than a single prompt. If that holds up outside Moonshot’s evaluations, the model’s importance will come from the length of the work it can carry, not the size printed on its spec sheet.

That distinction also makes the launch easier to judge. There are three questions worth asking: how much of the model is active for each token, what the long context enables in practice, and how much of the claimed performance can be verified today.

A huge model that uses a small fraction of itself at once

Kimi K3 is a mixture-of-experts model. It contains 896 experts but, according to Moonshot, activates 16 of them for a token. The full parameter count describes the model’s total capacity; it does not mean every token passes through all 2.8 trillion parameters.

That sparse design is central to the model, not an implementation footnote. A router selects a small set of experts for the current input, allowing the model to grow its total capacity without paying the dense-compute cost of using the entire network on every step. The hard part is keeping routing stable and preventing a few experts from receiving a disproportionate share of the work.

Moonshot says K3 addresses that problem with a framework it calls Stable LatentMoE and a balancing method derived from router-score quantiles. The model also uses Kimi Delta Attention and Attention Residuals, two architectural changes intended to improve information flow across long sequences and deep networks. Together with changes to training and data, the company claims these choices deliver roughly 2.5 times the scaling efficiency of Kimi K2.

That last figure is a company-reported result. The technical report has not been published yet, so there is not enough detail to independently examine what the efficiency comparison includes. Still, the architecture reveals the real strategy: Moonshot is trying to make extreme parameter scale usable rather than merely achievable.

One million tokens matters only if the model can use them

Kimi K3 supports a one-million-token context window. A window that large can hold a substantial codebase, a collection of research material, a long tool history, or some combination of all three.

Capacity alone does not produce a capable agent. The model has to recover the right detail after hundreds of intermediate steps, keep its plan coherent, interpret tool results, and avoid compounding an early mistake. A long context can become an expensive archive if the model cannot decide what matters inside it.

Moonshot’s examples are designed to show more than retrieval. In one reported 48-hour run, K3 used open-source electronic-design tools to design and verify a chip for a small model. In another case, it worked through more than 20 papers, evaluated over 300 equations of state, and produced more than 3,000 lines of Python for a computational astrophysics workflow. The company also says the model built a compact Triton-like GPU compiler with its own intermediate representation, optimization passes, and PTX generation pipeline.

These are unusually ambitious demonstrations because each task creates many chances for a model to drift. A compiler that performs well on an isolated kernel is useful; a compiler that supports an end-to-end training run tests whether all the pieces remain compatible. That is the kind of continuity a long-horizon coding model needs.

The examples remain vendor case studies, however. They show what Moonshot chose to test under conditions it controlled. They do not tell us how often comparable attempts fail, how much scaffolding the runs required, or how reliably another team could reproduce the result.

The product surface is part of the model story

Moonshot is putting K3 into Kimi’s web app, Kimi Work, Kimi Code, and its API. That distribution matters because agent performance depends on more than model weights. The harness decides which tools are available, what history is returned to the model, how failures are handled, and when a task should stop.

K3 makes that dependency unusually visible. Moonshot warns that the model can become unstable if an agent harness does not preserve its full thinking history, or if a session switches to K3 midway through. The company recommends a verified compatible harness such as Kimi Code. It also notes that the model may act too proactively when instructions are ambiguous, which means developers may need tighter system prompts or AGENTS.md constraints.

Those limitations are not minor for production agents. A model that continues working for longer also has more time to make an unauthorized decision. Endurance increases the value of good tool permissions, checkpoints, observability, and stop conditions.

The 1M-token window should therefore be understood as one component of a system. A useful long-horizon agent needs the model, the context-management strategy, the tool layer, and the control surface to work together.

“Open” is a direction today, not the finished state

Moonshot calls Kimi K3 the first open model in the three-trillion-parameter class. As of the launch announcement, the model is available through Kimi’s products and API, while the full weights are scheduled for release by July 27, 2026. The company says a technical report with more detail on architecture, training, and evaluation will arrive alongside further materials.

That timing matters. Until the weights, license terms, inference guidance, and technical report are available, developers cannot fully evaluate what “open” permits or how practical independent deployment will be. Moonshot recommends supernode configurations with at least 64 accelerators for inference, which also puts self-hosting far beyond an ordinary workstation or modest server cluster.

Moonshot is candid about another boundary: it says K3 still trails the strongest proprietary models overall and has a noticeable user-experience gap. Its benchmark results also mix different agent harnesses and include internal knowledge-work evaluations. That does not invalidate them, but it makes direct ranking less clean than a single chart suggests.

The right posture is interest without premature certainty. K3 has a technically interesting sparse architecture, an unusually large context window, and case studies aimed at work that lasts far longer than a benchmark prompt. Independent testing will decide whether those pieces form a dependable agent or an impressive launch demonstration.

The parameter count is the least actionable number

Kimi K3’s 2.8 trillion parameters establish scale. Its selective activation makes that scale computationally plausible. The one-million-token context gives it room to retain a long trajectory. None of those features, by itself, proves that the model can finish difficult work reliably.

The better test is operational: can K3 preserve the right state, recover from mistakes, respect boundaries, and produce a result that survives external verification after hours of tool use?

That is what long-horizon intelligence means in practice. The launch makes a serious case for it. The weights, technical report, and independent runs will show how much of that case travels beyond Moonshot’s own stack.