Moonshot's frontier model — 2.8 trillion parameters, a million-token window, and thinking effort you can dial up when the task deserves it.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Kimi K3 is Moonshot AI's frontier model, and the specification that matters most is not the parameter count. It is the output ceiling.
K3 pairs a 1,048,576-token context window with a 131,072-token maximum output. Plenty of models now read a million tokens. Far fewer will write back more than sixty-odd thousand. If your work ends in a long artefact — a migration plan, a full test suite, a translated codebase — that second number is the one that decides whether the model can finish the job in one pass.
K3 is built for long-horizon coding, knowledge work and deep reasoning, and it supports toggleable max-effort thinking.
That toggle is the interesting design choice. Reasoning models usually make you pick a tier and live with it: fast and shallow, or slow and thorough, chosen at the point you select the model. K3 moves the decision to the request. Bulk classification runs cheap; the one architectural question that actually matters gets full deliberation. For an agent that spends most of its loop on routine steps and occasionally hits something hard, that is the right shape.
K3 runs near $3.00 per million input tokens and $15.00 per million output tokens.
Read that against the output ceiling before you celebrate. A model that can emit 131,072 tokens in one response can also bill you for 131,072 tokens in one response. Long-horizon agent runs are exactly the workload where costs compound quietly, so cap output length in code rather than trusting the prompt to keep answers short.
K3's licensing is reported inconsistently. Some registries list its weights as open; Moonshot's own materials describe the model's capabilities without making an open-weight claim. The previous generation shipped under a permissive licence with downloadable weights, which is probably where the confusion starts.
If your plan depends on self-hosting rather than calling an API, confirm the licence directly with Moonshot before you build around it.
K3 is a considered choice, not a default. For high-volume, low-stakes calls it is more model than the task needs, and something like Gemini 2.5 Flash will serve the same request for a fraction of the cost. Where K3 earns its rate is sustained agent work — long runs, many tool calls, a large artefact at the end.
| Field | Value |
|---|---|
| Provider | Moonshot AI |
| Parameters | 2.8 trillion |
| Context window | 1,048,576 tokens |
| Maximum output | 131,072 tokens |
| Modality | Natively multimodal |
| Reasoning | Supported, effort is toggleable |
| Tool calling | Supported |
| Structured output | Supported |
| Released | 16 July 2026 |
Pricing and availability shift as providers adjust their rates. Confirm your own before you budget.
Compare it against the rest of the AI model directory.
Use the recorded facts as a shortlist, then validate the model against representative inputs and production constraints.
Kimi K3 is categorized for Any-to-any, Image-text-to-text, Image-to-text, Text generation, Text-to-text AI models, Video-text-to-text. Its current record accepts text, image, video and produces text. Confirm file formats, preprocessing, and provider-specific request schemas before implementation.
The directory records 1.0M context and 131.1K maximum output. Listed token prices are $3.00 input and $15.00 output per 1M tokens. Treat missing values as unknown and recheck current commercial terms.
Recorded output speed is — and time to first token is —. Hosting route, region, prompt length, concurrency, and provider load can materially change both measurements.
This page separates sourced model facts from incomplete fields. Benchmark evidence is displayed only when its source is linked and verified. Before choosing Kimi K3, test task quality, tool reliability, safety behavior, data controls, rate limits, and total cost on the exact route you intend to use.
Related approved models in Any-to-any. Compare specifications and verify fit for your workload.
Meta
Muse Spark is a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration.
OpenAI
Cost-efficient GPT-5.6 model for fast, high-volume workloads
OpenAI
Frontier GPT-5.6 model for complex professional work, coding, and agentic workflows
OpenAI
OpenAI's mid-tier GPT-5.6 reasoning model with a 1.05M token context window, function calling, and vision, priced between GPT-5.5 and GPT-5.5 Pro.
OpenAI
Default frontier GPT for coding, computer use, research, and knowledge work