Kimi K3 Reaches Frontier Scores With 2.8T Parameters and 1M Context
Moonshot AI's Kimi K3 scores 57 on the current Artificial Analysis Intelligence Index, ahead of Claude Opus 4.8, GPT-5.5, and DeepSeek V4 Pro. Its full weights are promised for July 27, and its API price is 30% of Claude Fable 5's, not one eighth.
Moonshot AI released Kimi K3 on July 16. The name is K3, not T3. It is a 2.8-trillion-parameter mixture-of-experts model with native vision, a 1-million-token context window, and a clear bias toward long coding and agent jobs.
The early numbers put it in an unusual spot. K3 is close to the best closed models on several current evaluations, but it also costs far more than DeepSeek V4. The open-weight claim needs a date attached to it too: Moonshot says the weights will arrive by July 27, 2026. They were not available when this article was published.
The benchmark picture
Artificial Analysis currently gives Kimi K3 a score of 57 on Intelligence Index v4.1. GPT-5.6 Sol at max effort scores 59 and Claude Fable 5 scores 60. K3 sits just behind them, while Claude Opus 4.8 scores 56, GPT-5.5 scores 55, and DeepSeek V4 Pro at max effort scores 44.
That supports the broad claim that K3 has moved past Opus 4.8, GPT-5.5, and DeepSeek V4 on this particular composite benchmark. It does not mean K3 wins every test. Artificial Analysis combines coding, scientific reasoning, knowledge, long-context work, and agent tasks into one score, so a one-point lead is closer to a photo finish than a knockout.
Moonshot's launch table is similarly mixed. K3 scores 88.3 on Terminal-Bench 2.1, half a point behind GPT-5.6 Sol and ahead of Opus 4.8, Fable 5, and GPT-5.5. It leads Program Bench with 77.8 and SWE Marathon with 42.0. On DeepSWE, though, GPT-5.6 Sol and Fable 5 remain ahead. The runs also use different agent harnesses, including Kimi Code, Claude Code, and Codex. That makes the table useful, but not perfectly controlled.
What 2.8 trillion parameters means here
K3 does not activate the whole model for every token. Its Stable LatentMoE design routes work through 16 of 896 experts. Moonshot has not yet published a simple active-parameter total, so any precise number circulating before the technical report is guesswork.
Two new pieces sit underneath that scale. Kimi Delta Attention is designed to make attention more efficient across long sequences. Attention Residuals let the network retrieve earlier representations across depth instead of repeatedly piling them together. Moonshot says this combination delivers about 2.5 times the scaling efficiency of Kimi K2.
The model accepts text and images and returns text through the API. Its context window reaches 1,048,576 tokens, and max reasoning effort is the default. Kimi Code also maps ultra, max, and xhigh to that max setting. Low and high modes are supported, but changing effort or switching models inside an old session can invalidate the prompt cache and hurt output quality.
This is not a model most people will run under a desk. Moonshot recommends supernode deployments with 64 or more accelerators and uses MXFP4 weights with MXFP8 activations. The open weights matter for researchers and large inference providers, not because a 2.8T model suddenly became a laptop download.
The features are broader than chat
Moonshot is selling K3 as a working agent rather than a better answer box. Its launch examples include a browser-based 3D game that iterates against screenshots, a GPU compiler built from scratch, and a 48-hour chip-design run using open-source EDA tools. Another example turns more than 20 astrophysics papers into a numerical pipeline, 3,000 lines of Python, and an interactive report.
Those are Moonshot case studies, not independent benchmarks. They do show where the model is aimed: software work that mixes code, terminal tools, images, documents, and repeated self-correction. Inside Kimi's own products, K3 also powers research reports, spreadsheets, slides, dashboards, and video-editing workflows. The public API currently documents text and image input with text output, so the product demos should not be read as a video-generation API claim.
The price story needs a correction
Kimi K3 costs $0.30 per million cached input tokens, $3 per million uncached input tokens, and $15 per million output tokens on Moonshot's API. Claude Fable 5 costs $10 for input and $50 for output. On those matching token rates, K3 costs 30% as much as Fable 5, which is about 70% cheaper.
It is not one eighth of the price. K3 is also 40% cheaper than Claude Opus 4.8 at its standard $5 input and $25 output rates. Claude Sonnet 5 is a different comparison again: its introductory price through August 31 is $2 input and $10 output, below K3. There is no single accurate price for a generic label such as "Claude 5."
DeepSeek keeps the cost advantage. DeepSeek V4 Pro is listed by Artificial Analysis at $0.435 per million input tokens and $0.87 per million output tokens. K3's independent benchmark score is higher, and it adds native image input, but its uncached input is almost seven times the price and its output is more than seventeen times the price. For high-volume text work, that gap can matter more than thirteen points on a composite index.
Open weights are promised, not shipped
Moonshot calls K3 the first open 3T-class model, but the launch post says the full weights will be released by July 27. It also says the technical report, training details, and fuller evaluation notes are still coming. Until the weights and license appear, K3 is an API model with a public open-weight commitment, not a model developers can already download and inspect.
That distinction should disappear if Moonshot ships on schedule. For now it is worth keeping. Artificial Analysis still labels K3 proprietary because the weights are not public on the article date.
Early limits worth reading
Moonshot lists two practical problems. First, K3 is sensitive to missing thinking history, so an agent that fails to replay the full history can become unstable. Second, it can be too proactive when the request is vague and may make decisions the user did not authorize. Moonshot recommends stronger system instructions or an AGENTS.md file for tightly bounded work.
The company also concedes that K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall user experience. That is a more useful description than declaring a universal winner. K3 has reached the frontier group on current tests. Whether it belongs in a production agent will depend on harness compatibility, token use, and how much supervision the job can tolerate.
Sources
Frequently asked
How many parameters does Kimi K3 have?
Kimi K3 has 2.8 trillion total parameters. It uses a mixture-of-experts design that routes each token through 16 of 896 experts, but Moonshot AI has not yet published a simple active-parameter total.
Is Kimi K3 open source?
Moonshot AI has committed to releasing the full model weights by July 27, 2026. The weights were not public when this article was published on July 18, so K3 is best described as an announced open-weight model rather than an already downloadable one.
Is Kimi K3 one eighth the price of Claude 5?
No. Kimi K3 costs $3 per million input tokens and $15 per million output tokens. Claude Fable 5 costs $10 and $50, making K3 about 70% cheaper. Claude Sonnet 5 has separate, lower introductory pricing, so the phrase Claude 5 is not precise enough for a price comparison.