What Is Kimi K3? Benchmarks, Pricing, Context Window, and Open-Weight Plans
Kimi K3 is Moonshot AI's new flagship model for long-horizon coding, agentic knowledge work, visual understanding, and deep reasoning. Announced on July 16, 2026, it combines a 2.8-trillion-parameter Mixture-of-Experts architecture with a one-million-token context window and an unusually sparse routing design.
The headline specifications are impressive, but the details matter. The strongest benchmark results currently come from Moonshot's own launch evaluation, API output is significantly more expensive than earlier Kimi models, and the promised downloadable weights were not yet available when this article was published on July 23, 2026. This guide separates what is already usable from what remains a release plan.
Source: kimi.com
Key takeaways
- Kimi K3 is a 2.8T-parameter multimodal MoE model that activates 16 of 896 routed experts for each token.
- The context window is 1,048,576 tokens, with automatic prefix caching and no higher API price tier for longer prompts.
- Official API pricing is $0.30 per million cache-hit input tokens, $3.00 per million uncached input tokens, and $15.00 per million output tokens.
- Moonshot reports leading or near-leading results on several coding, browsing, spreadsheet, automation, and visual-agent benchmarks, although the evaluation setups are not identical across models.
- The weights were scheduled for release by July 27, 2026. As of July 23, the checkpoint, final license, file sizes, and practical community deployment instructions were not yet available for inspection.
What is Kimi K3?
Kimi K3 is the successor to Moonshot AI's Kimi K2 generation and the company's largest model so far. Moonshot describes it as the first open model in the three-trillion-parameter class. It is designed less as a lightweight chat model and more as an always-reasoning agent that can maintain state across long engineering, research, document, and tool-use workflows.
The model is natively multimodal. It accepts text, images, and video through supported Kimi products and API input formats, then produces text and tool calls. Moonshot positions K3 around three main workloads: long-horizon software engineering, end-to-end knowledge work, and tasks that combine visual feedback with code or structured tools.
Kimi K3 is available through Kimi.com, the Kimi mobile applications, Kimi Work, Kimi Code, and the Kimi API. Developers use the API model ID kimi-k3. The API documentation states that flagship access is unlocked after a successful account top-up of at least $1, with rate limits determined by the account tier.
Readers who need broader background on the product family can start with Zerlo's guide to Kimi AI and Moonshot AI. K3 is the model release; Kimi is the wider assistant, application, and developer ecosystem.
Kimi K3 architecture: why 2.8 trillion parameters do not mean 2.8 trillion active parameters
Kimi K3 uses a sparse Mixture-of-Experts design. The network contains 896 routed experts, but only 16 are selected for each token. This allows the model to store a very large amount of specialized capacity without running every expert on every forward pass. The full checkpoint still has to be stored and distributed across an inference cluster, so sparse activation reduces computation more than it reduces the storage challenge.
Moonshot highlights four architectural components:
- Kimi Delta Attention (KDA): a linear-attention mechanism used in three of every four attention layers to make long-sequence processing more efficient.
- Gated Multi-Head Latent Attention: the fourth layer in the cycle retains full global attention for precise information retrieval.
- Attention Residuals: blocks can selectively retrieve representations from earlier depths instead of relying only on a single continuously accumulated residual stream.
- Stable LatentMoE: the expert-routing framework that supports the extremely sparse 16-of-896 expert configuration.
The company says these changes, combined with updated training methods and data recipes, provide roughly 2.5 times the overall scaling efficiency of Kimi K2. That figure is a Moonshot claim rather than an independently reproduced measurement. K3 also uses quantization-aware training with MXFP4 weights and MXFP8 activations, and Moonshot recommends supernode deployments with at least 64 accelerators.
A simple storage calculation illustrates the scale. At four bits per parameter, 2.8 trillion parameters represent about 1.4 terabytes of raw weight data before metadata, routing structures, runtime buffers, KV cache, redundancy, and framework overhead. This is why an eventual open-weight release will be valuable for researchers and hosting companies but will not make Kimi K3 a normal desktop model.
Kimi K3 benchmarks
Moonshot's launch materials compare Kimi K3 with GPT-5.6 Sol, GPT-5.5, Claude Fable 5, Claude Opus 4.8, and GLM-5.2. The company's own summary is relatively restrained: K3 still trails the most capable proprietary systems overall, but it reaches frontier-level performance and leads many individual tasks.
| Benchmark | Kimi K3 score | Position in Moonshot's launch chart |
|---|---|---|
| DeepSWE | 67.5 | Behind GPT-5.6 Sol and Claude Fable 5 |
| Terminal Bench 2.1 | 88.3 | Second, 0.5 points behind GPT-5.6 Sol |
| Program Bench | 77.8 | Highest reported score in the chart |
| SWE Marathon | 42.0 | Highest reported score in the chart |
| FrontierSWE | 81.2 | Second behind Claude Fable 5 |
| BrowseComp | 91.2 | Highest reported score in the chart |
| SpreadsheetBench 2 | 34.8 | Highest by 0.1 point over Claude Fable 5 |
| AutomationBench | 30.8 | Highest reported score in the chart |
Source: kimi.com
Moonshot reports particularly strong coding-agent results. K3 leads Program Bench and SWE Marathon in the launch chart while placing close to the top on Terminal Bench 2.1 and FrontierSWE.
How to read the benchmark claims
These numbers should not be treated as one perfectly controlled league table. Moonshot used Kimi Code, Claude Code, or Codex harnesses depending on the benchmark. Some competitor results came from official leaderboards or release posts, while other models were rerun by Moonshot. The launch table also notes possible fallback behavior for Claude Fable 5 and guardrail interruptions for GPT-5.6 Sol on some tasks.
All headline K3 results were obtained at maximum reasoning effort. That setting can improve quality but may also increase latency and output-token consumption. Several evaluations are internal, including Kimi Code Bench 2.0 and parts of the knowledge-work assessment. Until the weights and evaluation tooling are public, third parties cannot fully reproduce the complete result set.
The most useful reading is therefore not “K3 beats every closed model.” A better conclusion is that K3 appears highly competitive for long-running coding and agent workflows, with standout first-party results on software engineering, browsing, automation, spreadsheets, and visual document tasks. Real product quality will still depend on instruction following, latency, tool integration, safety behavior, and stability over long sessions.
Source: kimi.com
K3 also posts strong launch-day results outside pure coding, including first place in the shown BrowseComp, AutomationBench, and SpreadsheetBench 2 comparisons.
Kimi K3 pricing
The official Kimi API uses three token rates. Prices exclude taxes and use one million decimal tokens as the billing unit.
| Token category | Price per 1M tokens | When it applies |
|---|---|---|
| Cache-hit input | $0.30 | A repeated prompt prefix is served from Kimi's automatic context cache |
| Cache-miss input | $3.00 | New or changed input that must be processed normally |
| Output | $15.00 | Generated reasoning and answer tokens billed as output |
K3 has flat per-token pricing across its context range. There is no separate surcharge once a request crosses a 200K or similar threshold. However, long prompts can still be expensive in absolute terms, and an always-thinking model can produce substantial output. A request with one million uncached input tokens would cost $3 before output; the same cached prefix would cost $0.30.
Automatic caching does not require a cache ID or a separate API setting. The documentation says the previous prompt must exceed 256 tokens, and later requests should preserve the long prefix unchanged to have a chance of hitting the cache. This makes K3 more economical for repeated analysis of the same repository, document collection, or knowledge base than for constantly changing one-off prompts.
How large is the Kimi K3 context window?
Kimi K3 supports 1,048,576 tokens of context. In practical terms, that can hold very large codebases, extensive document collections, long agent histories, or combinations of text and visual inputs. The exact number of words, pages, or files varies because tokenization depends on language, formatting, source code, and embedded data.
The API documentation lists a default max_completion_tokens value of 131,072 and allows it to be raised as high as 1,048,576. That upper limit should not be interpreted as a recommendation to generate million-token answers. It is primarily an architectural ceiling for unusually long workflows.
A large context window also does not guarantee perfect recall. Moonshot explicitly warns that K3 was trained to preserve its thinking history. Agent frameworks should pass the complete assistant message back into subsequent turns. Dropping reasoning history, switching models in the middle of a session, or using an incompatible harness can make generation quality unstable.
Is Kimi K3 open source or open weight?
Moonshot calls Kimi K3 an “open” and “open-source” three-trillion-class model, but timing is crucial. The company announced that the full weights would be released by July 27, 2026, together with more technical details. This article was published four days before that deadline.
As of July 23, K3 was usable through Kimi products and the official API, but the full checkpoint was not yet listed on Moonshot's public Hugging Face organization page. The final model license, weight shards, checksums, exact storage footprint, supported inference engines, and community-tested deployment procedures therefore could not yet be verified.
The precise description at publication time is a proprietary hosted model with a promised open-weight release. After the checkpoint appears, the license will determine whether commercial use, modification, redistribution, and hosted services are permitted. Open weights do not automatically mean that the training data, complete training code, or reproducible training pipeline will also be published.
This distinction is covered in more detail in Zerlo's overview of China's open-weight AI ecosystem. K3 is also a useful scale comparison with GLM-5 and its open-weight agentic coding approach.
Where can you use Kimi K3?
- Kimi.com and mobile apps: consumer access through the web and the current iOS, Android, and HarmonyOS applications.
- Kimi Work: the desktop knowledge-work environment, with K3 support in version 3.1.0 or later for Windows and Apple silicon Macs.
- Kimi Code: terminal-based coding access, where users select K3 through the model menu.
- Kimi API: OpenAI-compatible developer access using the
kimi-k3model identifier. - Kimi Enterprise: organization accounts with enterprise privacy and member-management features.
- Self-hosting: planned after the weight release, but realistic deployment will require large multi-accelerator infrastructure.
Who should consider Kimi K3?
K3 is most compelling for teams that need an agent to operate across unusually large working sets: complex repositories, long technical histories, large PDF collections, data-heavy research, visual software workflows, or repeated queries over a stable knowledge base. Its cache pricing is designed to reward that repeated-prefix pattern.
It is less obviously suited to short, inexpensive chat completions. The model always reasons, output costs $15 per million tokens, and maximum-effort execution can be slower than a lightweight non-reasoning model. Teams should compare total task cost rather than input price alone: retries, long reasoning traces, tool calls, and agent runtime can matter more than the headline token rate.
Organizations interested primarily in open deployment should wait for the actual weight and license release before making infrastructure or compliance decisions. Even if the checkpoint is permissively licensed, a 2.8T model is likely to be consumed through specialized inference providers rather than downloaded onto ordinary workstations.
Limitations and open questions
- Independent verification is incomplete: the full checkpoint was not public at the publication date.
- Benchmark setups vary: different harnesses and cited third-party scores limit direct comparability.
- Always-on reasoning can increase cost and latency: lower effort may help, but it is not equivalent to a non-thinking mode.
- Long-session stability depends on history handling: Moonshot warns against dropping thinking history or switching models mid-session.
- The model can be overly proactive: the launch documentation recommends explicit behavioral limits for applications that must not improvise beyond narrow boundaries.
- Self-hosting is a cluster-scale project: sparse activation does not remove the need to store and distribute the enormous checkpoint.
FAQ
What is Kimi K3?
Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship multimodal model. It is designed for long-horizon coding, research, tool use, visual reasoning, and knowledge work, with a one-million-token context window.
How much does the Kimi K3 API cost?
The official rate is $0.30 per million cache-hit input tokens, $3.00 per million uncached input tokens, and $15.00 per million output tokens, excluding applicable taxes.
What is the Kimi K3 context window?
Kimi K3 supports 1,048,576 tokens of context. The API defaults to a maximum completion limit of 131,072 tokens, which can be raised to 1,048,576 for specialized workloads.
Is Kimi K3 available on Hugging Face?
Not yet at the time of publication on July 23, 2026. Moonshot said the full weights would be released by July 27. The official Hugging Face organization page should be checked again after that date.
Can Kimi K3 run locally?
It is not a practical local model for consumer hardware. A rough four-bit weight calculation alone is about 1.4 TB before runtime overhead, and Moonshot recommends deployment configurations with at least 64 accelerators.
Is Kimi K3 better than GPT or Claude?
No universal winner can be established from the launch data. K3 leads several reported coding and agent benchmarks, while Moonshot itself says the model still trails the strongest proprietary systems overall. Independent testing after the weight release will be more informative.
Does Kimi K3 support images and video?
Yes. K3 has native visual understanding and supports image and video input through documented Kimi products and API formats. The API requires supported uploaded-file or encoded input formats rather than arbitrary public image URLs.
Bottom line
Kimi K3 is one of the most ambitious AI model launches of 2026: a 2.8T sparse MoE with native vision, a one-million-token context window, strong first-party coding and agent benchmarks, and competitive cached-input pricing. It is already available as a hosted model, but the open-weight story was still a promise on July 23 rather than a downloadable, licensed checkpoint.
Developers can evaluate the API now, especially for repeated long-context workflows that benefit from automatic caching. Researchers and infrastructure teams should wait for the July 27 weight release, then verify the license, model files, supported serving stacks, and independent benchmark results before treating K3 as a production-ready open-weight deployment.