Kimi K3 is Moonshot AI's 2.8-trillion-parameter flagship model for long-horizon coding, knowledge work, visual reasoning, and tool-using agents. It combines a 1-million-token context window with native image and video understanding, always-on reasoning, structured output, and an unusually sparse Mixture-of-Experts architecture.
The headline numbers are impressive, but they do not answer the practical questions. Is Kimi K3 better than the strongest closed models? Is its 1M context useful or merely expensive? Can developers self-host a 2.8T model? And what limitations matter when K3 controls tools for hours rather than answering one prompt?
This review separates confirmed capabilities from launch claims and explains where Kimi K3 fits in a production model stack.
Poyo availability updated July 22, 2026. Kimi K3 is available through Poyo.ai as
kimi-k3and through Moonshot AI's official products. Moonshot says the complete model weights are scheduled for release by July 27, 2026; this article will be updated when the weights and license are public.
Kimi K3 at a glance
| Specification | Kimi K3 |
|---|---|
| Provider | Moonshot AI |
| Model ID | kimi-k3 on the official Moonshot API |
| Total parameters | 2.8 trillion |
| Architecture | KDA, Attention Residuals, Stable LatentMoE |
| Experts | 896 total, 16 activated |
| Context window | 1 million tokens |
| Maximum completion | 131,072 tokens by default; configurable up to 1,048,576 |
| Inputs | Text, images, and uploaded video |
| Output | Text |
| Reasoning | Always on; low, high, and max effort |
| Agent features | Tool calling, required tool choice, dynamic tool loading |
| Structured output | Strict JSON Schema |
| Context caching | Automatic |
| Official API status | Available |
| Poyo.ai status | Available as kimi-k3 |
| Full weights | Announced for release by July 27, 2026 |
Kimi K3 uses extreme scale differently from a dense 2.8T model. Stable LatentMoE routes each token through 16 of 896 experts, so total parameter count is not the same as active compute on every token. That distinction matters when comparing architecture, intelligence, and deployment cost.
What Kimi K3 is designed to do
Moonshot positions K3 around two broad workloads: long-horizon software engineering and end-to-end knowledge work.
Long-horizon coding
K3 is built to remain productive across long engineering sessions. The official examples emphasize large-repository navigation, terminal orchestration, GPU kernel optimization, compiler development, frontend engineering, game development, CAD, and scientific coding.
This is a different target from autocomplete. A long-horizon coding model must repeatedly inspect state, choose a tool, interpret the result, revise its plan, and verify the final artifact. Persistence and recovery can matter more than the quality of the first code sample.
Vision in the coding loop
K3 has native visual understanding. It can use screenshots and rendered output as feedback while changing code. That supports frontend iteration, visual QA, games, 3D scenes, diagrams, and other workflows where passing unit tests is not enough.
The official API accepts image data and uploaded video. Public image URLs are not accepted directly in vision messages: images must be supplied as base64 or through supported file handling, and video is referenced through a Moonshot file ID.
Knowledge work
K3 is also aimed at research, reports, spreadsheets, slides, dashboards, and multi-stage analysis. A 1M-token context can hold a large evidence set, but useful knowledge work still requires source control, citation checks, and verification. A large window reduces the need to discard context; it does not guarantee that every fact will be recalled or weighted correctly.
How good are the Kimi K3 benchmarks?
Kimi's launch report covers repository repair, terminal tasks, whole-program construction, knowledge work, automation, browsing, and multimodal evaluations. The reported pattern is more important than any single score: K3 is strongest on long, agentic workflows and remains competitive with leading closed models across the official suite.
The caveats matter:
- Kimi K3 results generally use maximum reasoning effort.
- Models were not always evaluated with the same agent harness.
- Some tasks ran on H20-calibrated environments rather than the benchmark's standard hardware.
- The official report notes fallback behavior for at least one competing model.
- Independent broad-intelligence rankings do not place K3 first overall.
Independent analysis currently places K3 near the frontier rather than clearly above it. That is consistent with Moonshot's own statement that K3 still trails the strongest proprietary models overall.
For the score tables, harness notes, and official-versus-independent distinction, read Kimi K3 Benchmarks Explained.
Kimi K3 API pricing
Moonshot lists flat pay-as-you-go rates with separate cached and uncached input prices.
| Token type | Official price per 1M tokens |
|---|---|
| Cache-hit input | $0.30 |
| Cache-miss input | $3.00 |
| Output | $15.00 |
There is no higher price tier simply because a request uses a long context. The total bill can still become large: one million uncached input tokens cost $3 before output, while 50,000 output tokens add $0.75.
Automatic caching can change the economics of repeated repository and research workflows. A stable 500K-token prefix costs $1.50 on a miss but $0.15 on a cache hit. Applications should log cache behavior rather than assuming every repeated request receives the lower rate.
See Kimi K3 API Pricing Explained for complete workload examples and caching strategies.
Kimi K3 vs GPT-5.6 Sol
K3 and GPT-5.6 Sol overlap in high-end coding, agents, reasoning, and professional work, but they represent different tradeoffs.
| Decision area | Kimi K3 | GPT-5.6 Sol |
|---|---|---|
| Primary advantage | 1M context and planned open weights | Stronger top-end closed-model performance |
| Coding | Strong long-horizon and visual coding | Strong general coding and terminal performance |
| Multimodal input | Native image and video understanding | Multimodal capabilities depend on current API surface |
| Context | 1M tokens | Verify the current endpoint limit |
| Open weights | Announced | No |
| Deployment | Official API or very large future self-hosted cluster | Hosted API |
| Production maturity | Newly launched | Established OpenAI platform |
No model wins every workload. K3 is compelling for teams that value long context, multimodal coding, and future weight access. GPT-5.6 Sol remains attractive when independent top-end quality, mature platform behavior, and closed hosted infrastructure matter more.
Read the full decision guide in Kimi K3 vs GPT-5.6 Sol.
Kimi K3 strengths
A genuinely large working context
One million tokens gives agents room for source code, documentation, tool history, visual evidence, and intermediate artifacts. This is particularly useful when aggressive retrieval would separate evidence that needs to be considered together.
Coding and vision in one model
K3 can inspect a rendered result instead of relying only on source code or text logs. That makes it relevant to UI engineering, games, CAD, data visualization, and media workflows.
Agent-oriented API features
The official API supports custom tools, forced tool use, dynamic tool loading, streaming, structured output, and long multi-turn sessions. These are production building blocks rather than chat-only features.
Open-weight direction
Moonshot describes K3 as the first open 3T-class model and has announced complete weights. If the release and license permit broad use, it will give infrastructure teams more control than closed API-only models. Availability, license terms, and community inference support must still be verified after release.
Kimi K3 limitations
Thinking history must be preserved
Moonshot warns that K3 was trained with preserved thinking history. A client that drops required assistant state, keeps only final text, or switches models in the middle of a session can cause unstable quality. This makes correct conversation-state handling essential.
It can be excessively proactive
Long-horizon training encourages the model to keep moving. In ambiguous situations, K3 may make decisions the user did not authorize. Production agents need explicit boundaries, scoped tools, approval steps, and clear instructions about when to stop.
The biggest model is not a local-laptop model
Even with MXFP4 weights and sparse expert activation, a 2.8T model is an infrastructure-scale deployment. Moonshot recommends supernode configurations with 64 or more accelerators. Open weights will not make K3 practical on a consumer GPU or ordinary workstation.
Output is relatively expensive
At $15 per million output tokens, verbose agent behavior can dominate the bill. Applications should set sensible completion limits and measure cost per verified task, not price per request.
Some release details are still pending
As of July 21, the full weights, final license, and technical report are not yet public. Articles claiming a complete local installation process before those artifacts exist should be treated cautiously.
Who should use Kimi K3?
K3 is a strong candidate for:
- repository-scale coding agents;
- terminal workflows that require persistence;
- frontend, game, CAD, and visual software iteration;
- long-document research and professional artifacts;
- multimodal analysis with images or uploaded video;
- teams evaluating future open-weight frontier deployments.
It may be unnecessarily large or expensive for:
- classification, extraction, and simple support;
- latency-sensitive short responses;
- small self-hosted environments;
- workflows that cannot preserve full conversation state;
- agents without strong permission and verification boundaries.
Kimi K3 verdict
Kimi K3 is one of the most consequential open-model launches of 2026 because it combines frontier-scale architecture, a 1M context window, native visual understanding, and a serious agent API. Its most convincing use case is not a one-shot benchmark question. It is sustained work across code, tools, documents, and visual feedback.
The model is not an automatic replacement for the strongest closed systems. Independent evaluation, output cost, state-management requirements, excessive proactiveness, and extreme self-hosting requirements all matter. The right production decision is to test K3 on representative long-running tasks and compare cost per verified result.
Follow the Kimi K3 model page for Poyo.ai availability, confirmed model IDs, and pricing.
Frequently asked questions
Is Kimi K3 open source?
Moonshot calls K3 an open 3T-class model and says complete weights will be released by July 27, 2026. As of July 21, developers should wait for the actual files and license before making deployment or commercial-use claims.
How large is the Kimi K3 context window?
Kimi K3 supports a 1-million-token context window. Automatic context caching is available on the official API.
Is Kimi K3 better than GPT-5.6 Sol?
Not universally. K3 is especially attractive for long context, planned open weights, multimodal coding, and agent workflows. GPT-5.6 Sol retains advantages in some independent and top-end evaluations. Test both on the same tasks.
Does Kimi K3 support images and video?
Yes. The official API supports image input and uploaded video input. It returns text rather than generated images or video.
How much does the Kimi K3 API cost?
Moonshot lists $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens.
Can Kimi K3 run locally?
Not on ordinary consumer hardware. The 2.8T model is intended for large distributed infrastructure, even though only a subset of experts is active per token.