Overview
Kimi K3 is Moonshot AI's July 2026 flagship model, following the coding-specialized K2.7 Code release with a broader flagship for long-horizon software work, knowledge tasks, reasoning, and visual inputs. The headline changes are the 2.8-trillion-parameter Mixture-of-Experts scale and 1-million-token context window; native vision continues a capability already present in earlier Kimi releases.
For teams evaluating a production model, the practical distinction is breadth. K2.7 Code targeted coding agents; K3 is available in Kimi, Kimi Work, Kimi Code, and the Kimi API for workflows that combine code, documents, images, research, and persistent agent tasks. It is live now, but Moonshot has not yet released its full weights or technical report.
What's New
A 1M-Token Flagship That Retains Native Vision
K3 raises the context window to one million tokens while retaining native visual input, which was already available in earlier Kimi models. That can reduce the need to split large repositories, long research corpora, or image-heavy tasks into separate model sessions. It does not remove the need for retrieval, context management, or evaluation on a real workload, but it changes the ceiling for tasks that need broad context in one run.
Moonshot describes the model as 2.8T parameters with 16 of 896 experts activated per token. Its Kimi Delta Attention and Attention Residuals architecture are implementation details rather than a buying criterion; the user-facing implication is a model intended for long-horizon coding, reasoning, and knowledge work rather than a narrow coding-only variant.
One Model Across Kimi's Product Surfaces
K3 is available in the Kimi web and app experience, Kimi Work, Kimi Code, and the API as kimi-k3. In Kimi Code, select it with /model. Kimi Work 3.1.0 or later adds K3 support along with Widgets and Dashboard, which Moonshot positions for persistent, visual knowledge-work outputs.
At launch, K3 uses maximum thinking effort by default. Moonshot says lower and higher effort modes will arrive later, so teams that need stable latency or spend controls should validate the current default before changing an existing production route.
Published Benchmarks Need Workload Validation
Moonshot reports strong results with maximum reasoning effort, including 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp, and 83.4 on MMMU-Pro with Python. These figures are useful signals, not universal guarantees: the official table uses different agent harnesses for different models, some comparison results use fallback behavior, and several benchmarks are internal. Test K3 with the same prompts, tools, context policy, and cost limits that your team uses today.
Availability & Access
K3 is available through Kimi, Kimi Work, Kimi Code, and the Kimi API. Kimi Work requires version 3.1.0 or later on Windows or Apple silicon Mac; API users select kimi-k3; and Kimi Code users can select the model with /model.
Rollout and Session Limitations
Moonshot says the full model weights will be released by July 27, 2026. The technical report is forthcoming, but the launch post does not assign it an explicit date. Until then, this is not a confirmed self-hosting option and no license, hardware, or local-inference recommendation should be inferred from K3's "open 3T-class" positioning.
Moonshot also warns that K3 was trained with preserved thinking history. If an agent harness does not return the required history, or switches an ongoing session from another model to K3, output quality can become unstable. Kimi Code is the compatibility path Moonshot explicitly recommends.
Pricing & Plans
Moonshot publishes API pricing for kimi-k3 per one million tokens:
| Usage | Price |
|---|---|
| Cache-hit input | $0.30 |
| Cache-miss input | $3.00 |
| Output | $15.00 |
K3 can also be accessed through Kimi product surfaces, but Moonshot's release announcement does not establish a K3-specific membership price or quota. API spend and subscription access should therefore be evaluated separately. The $0.30 entry price only applies when input is served from cache; new or uncached context is billed at the higher input rate.
Best For
- Engineering teams testing a long-context model for repository-scale coding, tool use, and visual debugging.
- Researchers and analysts whose work combines large document sets, charts, images, and iterative investigation.
- Kimi Code users who want to evaluate a general flagship model instead of the coding-specialized Kimi K2.7 Code.
- API teams that can benefit from cached context and are able to preserve reasoning history across multi-step tool calls.
- Product teams comparing agentic knowledge-work models with a defined evaluation harness and cost budget.
FAQ
Is Kimi K3 available now?
Yes. Moonshot says K3 is available in Kimi, Kimi Work, Kimi Code, and the Kimi API. Kimi Work needs version 3.1.0 or later, and API users select kimi-k3.
Is Kimi K3 open source or available for self-hosting?
Not yet on the evidence available at launch. Moonshot calls K3 an open 3T-class model and says its full weights will be released by July 27, 2026; the technical report is also forthcoming, but no explicit publication date is stated. Wait for the official weights, license, and deployment guidance before planning a local deployment.
What does Kimi K3 cost through the API?
The official API price is $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens, and $15.00 per million output tokens. Subscription access and API billing are separate questions.
How does K3 compare with K2.7 Code?
K2.7 Code was a coding-specialized release. K3 is Moonshot's broader flagship, expanding the maximum context from K2.7 Code's 256K to up to 1M while retaining multimodal input for coding, knowledge work, reasoning, and visual tasks. Moonshot has not published a like-for-like K3 versus K2.7 Code comparison for every workload, so test both for your own use case.
Should an existing agent switch to K3 mid-session?
Moonshot advises against it. K3 depends on preserved thinking history, and switching an ongoing session from another model or dropping required history can make output unstable. Start a compatible K3 session and validate the harness first.
