Kimi K3 is available on Novita AI as Moonshot AI’s flagship model, with 2.8 trillion total parameters, a 1,048,576-token context window, and text, image, and video input. Choose K3 when your application needs very large context or multimodal reasoning; choose a Kimi K2 variant when lower per-token cost, a smaller context window, or an established K2-specific integration matters more. Novita’s current K3 pricing is $3 per million input tokens, $15 per million output tokens, and $0.30 per million cache-read tokens, checked July 22, 2026.
Key Takeaways
- Kimi K3 is live on Novita AI under the exact model ID
moonshotai/kimi-k3. - The model page lists a 1M-token context and 1M maximum output, plus text, image, and video input capabilities.
- Current serverless pricing is $3/M input, $15/M output, and $0.30/M cache reads. Pricing can change, so confirm it on the model page before budgeting.
- K3 is the better Kimi choice for large repositories, long research packets, multimodal knowledge work, and workflows that need the newest flagship.
- K2.5 or K2 Instruct can be the more sensible choice for cost-sensitive, text-only, or already-integrated workloads.
What Is Kimi K3?
Kimi K3 is Moonshot AI’s latest Kimi model and the most capable model listed in Novita AI’s Kimi lineup. Its headline difference is scale: the Novita model page describes a 2.8-trillion-parameter model built with Kimi Delta Attention and Attention Residuals, with native visual understanding and a 1M-token context window.
That context size changes the practical question from “Can the model answer this prompt?” to “Can I give the model the full working set?” A coding agent can keep more of a repository in context. A research assistant can work from a large collection of documents without aggressive chunking. A multimodal workflow can combine text with images or video instead of splitting those inputs across separate steps.
The trade-off is cost. K3’s output price is materially higher than the prices listed for older Kimi endpoints on Novita, so a larger context window is most useful when it removes real application complexity or improves answer quality. It is not a reason to send every short chat request to the flagship.
Kimi K3 Specs and Pricing
The following values come from the Kimi K3 model page on Novita AI, checked July 22, 2026.
| Field | Kimi K3 details |
|---|---|
| Publisher | Moonshot AI |
| Model ID | moonshotai/kimi-k3 |
| Architecture note | 2.8T parameters; Kimi Delta Attention and Attention Residuals |
| Context window | 1,048,576 tokens (1M) |
| Maximum output | 1,048,576 tokens |
| Input | Text, image, video |
| Output | Text |
| Input price | $3 per million tokens |
| Cache-read price | $0.30 per million tokens |
| Output price | $15 per million tokens |
| Novita availability | Serverless API |
| API styles | OpenAI-compatible and Anthropic-compatible endpoints |
The pricing table is usage-based: you pay for tokens processed rather than provisioning a dedicated GPU. For a rough estimate, 10 million input tokens and 1 million output tokens would cost $45 before any other account-level charges or changes to the listed rates. Your own traffic mix matters more than the headline output price, especially if prompts are long and responses are short.
K3 also lists cache-read pricing. Reusing stable system instructions, tool definitions, or large reference material can reduce the cost of repeated requests when your application and request pattern support caching. Treat cache behavior as an optimization to measure, not as a guarantee that every repeated prompt will be charged at the cache-read rate.
Kimi K3 vs the Kimi K2 Family
K2 is not one single endpoint. The family includes older K2 Instruct variants and newer K2.5 and K2.6 releases, each with its own capability and pricing profile. The right comparison is therefore workload-based rather than a blanket claim that every K3 response is better.
| Decision factor | Kimi K3 | Kimi K2 family |
|---|---|---|
| Positioning | Newest flagship in the Novita Kimi lineup | Earlier variants with different cost and capability trade-offs |
| Context | 1M tokens | Varies by model; verify the selected endpoint |
| Input modalities | Text, image, video | Varies by model; K2.5 supports multimodal input |
| Current Novita K3 price | $3/M input, $15/M output, $0.30/M cache read | Varies by version; check the model page before comparing |
| Best reason to choose it | Very large working sets and frontier multimodal workflows | Lower cost, known behavior, or an existing integration |
| Migration risk | New model ID and new production baseline | Existing prompts, evaluations, and tool schemas may already be tuned |
K3 Compared With K2.5
K2.5 remains a reasonable choice for multimodal applications that do not need a million-token context. It can handle text, images, and video, and its Novita model page lists a smaller context window and lower current token rates than K3. Use K2.5 when your prompts fit comfortably within its limits and your priority is keeping inference spend predictable.
Move to K3 when the larger context removes a real bottleneck: repeated retrieval, repository summarization across many files, long video or document analysis, or agent runs that lose important state during compaction. If the workload is mostly short conversations or ordinary coding questions, K2.5 may deliver a better cost-to-latency trade-off.
K3 Compared With K2 Instruct
K2 Instruct is the simpler value option for text-first chat and agentic tool use. Its older Novita post lists a 128K context window and lower pricing than K3. That can still be enough for support agents, structured extraction, coding assistance on small repositories, and applications that already rely on K2 Instruct’s prompt behavior.
K3 is a better fit when 128K is not enough, when visual inputs are part of the core workflow, or when you want to evaluate Moonshot AI’s current flagship rather than preserve an older baseline. Do not migrate solely because K3 has more parameters: test your prompts, tool calls, output format, and latency with representative traffic.
When to Use Kimi K3
K3 earns its higher price when one or more of these conditions apply:
Large repositories and long-running coding agents
Use K3 for tasks that require a broad view of a codebase: cross-package refactors, architecture mapping, dependency migrations, or debugging that spans logs, tests, and implementation files. A 1M-token context does not automatically make an agent reliable, but it gives your orchestration layer more room before it must summarize or discard state.
Multimodal knowledge work
K3 accepts text, image, and video input. That makes it a candidate for workflows such as reviewing product screenshots with requirements, analyzing recorded demonstrations alongside documentation, or extracting decisions from mixed media. Keep the output contract narrow and validate extracted fields if the result feeds an automated system.
Long research and document synthesis
When a question depends on many primary documents, K3 can reduce the amount of retrieval and stitching code required to assemble a working context. This is most useful for comparison, synthesis, and cross-document questions; it does not remove the need to identify authoritative sources or check citations.
Flagship evaluation
If you are evaluating the current Kimi generation for a new product, K3 is the natural first endpoint to benchmark. Start with a small, representative test set and compare quality, latency, output length, and total cost against K2.5 or K2 Instruct before changing your production default.
When K2 Is the Better Choice
K3 is not the default answer for every request. Stay with a K2 endpoint when:
- Your prompts are short and text-only, so the 1M context provides no practical benefit.
- Your application is highly price-sensitive and output tokens dominate the bill.
- You already have a stable K2 prompt, tool schema, or structured-output evaluator in production.
- You need a known latency and have not yet tested K3 under your rate and concurrency profile.
- Your workload fits the selected K2 context window without frequent truncation or summarization.
A useful routing pattern is to keep K2.5 or K2 Instruct as the fast path and send only large-context, multimodal, or difficult tasks to K3. That lets you evaluate K3 on the requests where its added context and capability can pay for themselves.
How to Access Kimi K3 on Novita AI
The Kimi K3 page includes Playground access, model documentation, and serverless API details. For an OpenAI-compatible client, use Novita’s OpenAI base URL and the exact model ID:
from openai import OpenAI
client = OpenAI(
api_key="<NOVITA_API_KEY>",
base_url="https://api.novita.ai/openai",
)
response = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[
{"role": "user", "content": "Summarize the key risks in this design."}
],
)
print(response.choices[0].message.content)Keep the first production test small. Record prompt tokens, completion tokens, cache usage, time to first token, total latency, and task success. Then repeat the same evaluation with the K2 endpoint you would otherwise use. The model page also lists Anthropic-compatible access, which can be useful if your existing application already uses that client shape.
Conclusion
Choose Kimi K3 on Novita AI when the job is genuinely context-heavy or multimodal and you want to evaluate Moonshot AI’s newest flagship through a serverless API. Its 1M-token context, visual input, and current flagship positioning make it a strong candidate for long-horizon coding agents, document-heavy research, and mixed-media workflows.
Choose K2.5 or K2 Instruct when the workload is smaller, text-first, cost-sensitive, or already tuned around a K2 endpoint. The best production setup may be a router rather than a single winner: reserve K3 for requests that need its larger working set, and keep a lower-cost K2 path for routine traffic.
FAQ
Is Kimi K3 available on Novita AI?
Yes. Novita lists Kimi K3 as a serverless model with the model ID moonshotai/kimi-k3. Availability and pricing can change, so confirm the live model page before deployment.
How much does Kimi K3 cost on Novita AI?
As checked July 22, 2026, the listed rates are $3 per million input tokens, $15 per million output tokens, and $0.30 per million cache-read tokens. Check the model page for the current rate before estimating production spend.
Does Kimi K3 support images and video?
Yes. The Novita model page lists text, image, and video as input modalities, with text output.
Should I migrate from Kimi K2.5 to K3?
Migrate when your workload benefits from K3’s 1M-token context, newer flagship positioning, or multimodal workflow. If K2.5 already handles your prompts and your priority is lower cost, benchmark before switching.
What is the Kimi K3 model ID?
Use moonshotai/kimi-k3 in the Novita API configuration.
