Novita AI Adds Xiaomi's MiMo V2.6 Pro With a 1M-Token Context Window

Specifications and pricing

Novita AI lists the model under the ID xiaomimimo/mimo-v2.6-pro and serves it through the base URL https://api.novita.ai/openai. The catalog describes a serverless chat model with reasoning, function calling and structured-output support. It accepts text, image, video and audio input and returns text.

• Context window: 1,048,576 tokens
• Maximum output: 131,072 tokens
• Input price: $4.35 per 1M tokens
• Cached input price: $0.036 per 1M tokens
• Output price: $8.70 per 1M tokens
• Endpoint families: chat completions, Anthropic, Responses

Novita gives a worked example: an uncached request with 10,000 input tokens and 2,000 output tokens would cost about $0.0609, made up of $0.0435 for input and $0.0174 for output. The company advises using the usage data returned by the API for accounting rather than estimating from character count. It also warns that prices, limits and availability can change, and recommends checking the model page and pricing before setting a production budget.

How developers can call it

Because the endpoint follows the OpenAI chat-completions shape, the OpenAI Python client works by changing the base URL, API key and model ID. Direct HTTP requests go to /openai/v1/chat/completions, with the model ID in the JSON body. Novita advises keeping the API key in an environment variable or secret manager, not in source control, browser code or a client-side bundle.

The guide also suggests starting with a modest max_tokens value. In Novita's words, the catalog maximum is an upper limit, not a recommendation for every request, and keeping output close to the task makes cost and latency easier to control.

For teams already using MiMo V2.5 on Novita AI, the first migration step is small: keep the same base URL, swap in the new model ID and run the existing prompt suite. Novita cautions against assuming that a larger context limit or an extra input modality automatically improves an application. It recommends comparing answer quality, tool-call arguments, structured-output validity, latency and token usage, and keeping the old model available for rollback.

Suggested uses and cautions

Novita positions the model for tasks that benefit from a large working context or multi-step reasoning. The examples it names include repository-scale coding assistants, research and document workflows, tool-using agents, multimodal analysis and structured business workflows. For short prompts and high-volume classification, it says a smaller model can be easier to budget and faster to operate.

The guide stresses several implementation points:

• Context is a budget. Large prompts still cost money and can raise latency, so only relevant material should be retrieved.
• Validate structured output. Applications should still parse the returned data and reject missing fields, unexpected values and invalid JSON before triggering an external action.
• Keep tool execution outside the model. Function calling describes the action the model wants to take but does not grant access to systems, so the application should authenticate and authorize calls itself.
• Measure caching. The cached-input rate is much lower than the regular input rate, but Novita says the effect should be verified from response usage fields rather than assumed.

What it means for developers

For developers, the news is mainly about access: a Xiaomi reasoning model with long context, tool support and multimodal input is now reachable through a familiar OpenAI-style interface. Novita says the strongest reason to choose it is the combination of long context, reasoning, tool support and multimodal input behind one hosted API, not the headline context size alone.

The company recommends testing on a small production slice first, checking tool-call accuracy, structured-output validity, multimodal quality, timeout behavior and cost. As it notes, catalog features show what the endpoint supports but do not replace an evaluation against an application's own failure cases.