Kimi K3 launches with 2.8 trillion parameters and a 1M-token context

Moonshot AI has introduced Kimi K3, its most capable model to date. The launch quickly escaped the usual circle of people tracking Chinese AI labs. K3 combines an unusually large architecture, a one-million-token context window, native visual understanding and results that place it close to GPT-5.6 Sol and Claude Fable 5 on selected tasks.

That does not mean Kimi has beaten every competitor. In its official Kimi K3 technical overview, Moonshot explicitly says the model still trails the strongest proprietary systems from OpenAI and Anthropic overall. K3 matters for a different reason: a model announced for an open-weight release is already winning selected coding and agent evaluations, while its hosted API costs less than the most expensive frontier models.

Kimi K3 at a glance

• announced on July 16, 2026,
• 2.8 trillion total parameters in a Mixture-of-Experts architecture,
• 16 activated experts out of 896,
• up to 1,048,576 tokens of context,
• native support for text, images and video,
• available through Kimi, Kimi Work, Kimi Code and API,
• full model weights promised by July 27, 2026.

A much larger Kimi, but not 2.8 trillion active parameters

The 2.8 trillion parameter headline needs context. Kimi K3 uses a Mixture-of-Experts architecture, so it does not run the whole network for every token. Moonshot built 896 experts and selects 16 of them during computation. This design increases total model capacity without increasing the cost of every response in direct proportion to the total parameter count.

K3's 2.8 trillion parameters therefore should not be compared directly with the parameter count of a dense model. The figure describes the complete system, not the part used simultaneously at each generation step. Moonshot has not yet published a complete active-parameter count. More technical detail is expected with the forthcoming report.

Two architectural changes sit at the centre of K3:

• Kimi Delta Attention (KDA) is designed to reduce the cost of processing very long sequences,
• Attention Residuals (AttnRes) let the model retrieve representations selectively across depth instead of accumulating them in a fixed way.

Moonshot combines these ideas with Stable LatentMoE, Gated MLA and quantization-aware training. The weights use MXFP4 and activations use MXFP8. The company says the full set of changes delivers roughly 2.5 times better scaling efficiency than Kimi K2. That is a vendor claim that can be examined more thoroughly once weights and technical documentation are public.

How Kimi K3 compares in coding benchmarks

Coding produced the most visible launch result. On release day, Kimi K3 ranked first in Frontend Code Arena, where developers choose between outputs without seeing the model names. As the Associated Press reported, K3 moved ahead of leading US systems in this specific frontend ranking.

Moonshot also published a much broader evaluation package. The table below selects several useful results from the launch material:

Benchmark	Kimi K3	GPT-5.6 Sol	Claude Fable 5	Claude Opus 4.8
Terminal-Bench 2.1	88.3	88.8	84.6	84.6
Program Bench	77.8	77.6	76.8	71.9
FrontierSWE	81.2	71.3	86.6	66.7
SWE Marathon	42.0	39.0	35.0	40.0
DeepSWE	67.5	73.0	70.0	59.0
Kimi Code Bench 2.0	72.9	64.8	76.9	71.7

The result is mixed, which makes it more informative than the simple claim that Kimi beat Claude. K3 narrowly leads GPT-5.6 Sol on Program Bench and takes first place on SWE Marathon. Sol leads DeepSWE, while Fable 5 holds a substantial advantage on FrontierSWE and Moonshot's own Kimi Code Bench 2.0.

These numbers make K3 a serious coding model. They do not establish it as the best choice for every repository or software team.

Why the benchmark table is not a simple league ranking

Modern agentic benchmarks measure more than the base model. Kimi K3 was paired with environments including Kimi Code, OpenAI models used Codex, and Claude models ran through Claude Code or other harnesses specified in Moonshot's methodology. Each system has its own tools, system instructions and context-management strategy.

Moonshot also states that all K3 results were obtained at maximum reasoning effort. This can improve completion rates, but it may increase output tokens, latency and the effective price of a task.

Data provenance matters too. Kimi Code Bench 2.0 is an internal Moonshot test. FrontierSWE, Terminal-Bench and independent Arena preferences provide useful external context, but none replaces an evaluation on a team's actual codebase. The published results are a strong reason to test K3, not a complete procurement decision.

One million tokens changes the type of work K3 can attempt

Kimi K3 supports up to 1,048,576 tokens of context. In principle, that allows an agent to receive a large repository, project documentation, historical decisions and visual material in one session. For coding agents, a larger window can reduce aggressive summarisation and help preserve relationships between distant files.

Long context is not free memory. Sending more material makes requests more expensive and can increase latency. The model still needs to identify the relevant detail inside a huge input. Indexing quality, prompt structure and caching remain just as important as the maximum token count.

K3 also has native visual capability. It can inspect screenshots and combine them with code changes, which is useful for interfaces, games and CAD workflows. Moonshot demonstrates video editing, presentations and interactive reports as well. These are vendor examples, so teams should test repeatability with their own assets and acceptance criteria.

Beyond code: reports, spreadsheets and extended research

Moonshot positions K3 as a model for knowledge work, not merely a successor to Kimi K2.7 Code. The company shows interactive research reports, literature analysis, complex spreadsheet work and multiple agents cooperating on one problem.

In the launch material, K3 scored 34.8 on SpreadsheetBench 2, just ahead of Fable 5 at 34.7. The one-tenth-point difference is less important than the wider lesson: the best model still completes only about one third of these difficult, multi-step spreadsheet workflows. These systems can accelerate analysis and produce drafts, but they are not reliable autonomous accountants or financial analysts.

K3 also led Moonshot's AutomationBench and BrowseComp comparisons, while Fable 5 remained stronger across many tests of broad knowledge, reasoning and perception. K3's profile is becoming clear: extended agentic work, terminal use, coding, research and large collections of source material.

Kimi K3 API pricing

The model is available on the official Kimi API platform under the identifier kimi-k3. Pricing per million tokens is:

Token type	Price per 1M tokens
cache-hit input	$0.30
cache-miss input	$3.00
output	$15.00

This is substantially more expensive than earlier Kimi releases, especially on output. It remains below many premium frontier offerings. Context caching is crucial: Moonshot says its Mooncake inference infrastructure reaches a cache-hit rate above 90 percent on coding workloads. Applications that repeatedly use the same repository or document set could therefore see a large difference between the $0.30 and $3 input rates.

Token pricing alone is not enough for comparison. K3 launches with maximum reasoning effort, so it may generate more tokens than a nominally more expensive model that reaches a result with a shorter trajectory. Cost per correct, completed task remains the better metric.

Where Kimi K3 is available

K3 is already running in several products:

1. Kimi on the web and in mobile apps,
2. Kimi Work version 3.1.0 or later,
3. Kimi Code, where users select K3 through the /model command,
4. the official API as kimi-k3.

The term "open-weight" needs a date attached to it. Moonshot has promised the full model weights by July 27, 2026, but they are not public at the time of this article. Their release will allow the community to inspect deployment requirements, run independent evaluations and assess the licence terms.

This will not be an easy model to run on a home workstation. Moonshot recommends supernode configurations with at least 64 accelerators. Even after the weight release, K3 will mainly be practical for infrastructure providers, large research groups and companies consuming hosted endpoints.

Limitations Moonshot acknowledges

Moonshot identifies three issues that directly affect users:

• K3 is sensitive to incomplete reasoning history and may become unstable if a session is switched from another model,
• training for long-horizon tasks can make it excessively proactive, causing it to make decisions that were not clearly requested,
• its overall user experience still trails Claude Fable 5 and GPT-5.6 Sol.

The first limitation makes agent-harness compatibility particularly important. The second calls for explicit boundaries in system prompts and files such as AGENTS.md. An agent with terminal access should not interpret ambiguity as permission to perform a risky operation.

Is Kimi K3 another DeepSeek moment?

The comparison is understandable: a Chinese model is approaching proprietary leaders and is scheduled to receive public weights. The circumstances are not identical. K3 costs more than earlier Moonshot models, and its enormous architecture makes self-hosting difficult. It is not a lightweight release that most developers can run locally.

The significance lies elsewhere. Kimi K3 shows that open models no longer have to compete only on price. They can lead difficult coding evaluations, process images and operate across a million-token context. Moonshot is also unusually direct about the remaining gap: it does not describe K3 as the unconditional best model in the world.

For developers, research teams and companies building agents, this is one of the most important releases of 2026. The responsible verdict is more measured: Kimi K3 looks like a frontier model on selected workloads, but the decisive test starts after the weight release and independent measurement of its cost, reliability and quality on real projects.

(Photo: Oleg Gospodarec / Unsplash, license.)