Novita AI adds MiMo V2.6 Flash with 1M-token context and low token prices

What Novita AI lists for MiMo V2.6 Flash

Novita AI describes MiMo V2.6 Flash as the cost-efficient variant in Xiaomi's MiMo V2.6 line. On the platform it runs under the model ID xiaomimimo/mimo-v2.6-flash and is accessed through a serverless API.

The model accepts text, image, video and audio as input and returns text. Novita lists a context window of 1,048,576 tokens (1M) and a maximum output of 131,072 tokens (128K). Listed features are reasoning, function calling and structured outputs. The catalog also shows streaming and prompt-caching capabilities.

Three endpoint families are reported: Chat Completions, Responses and an Anthropic-compatible one. Novita says this lets applications pick the interface that fits their existing client and integration pattern.

Pricing and the role of caching

Novita says it checked the following prices against its model catalog on September 30, 2026:

• Input tokens: $0.14 per 1M tokens
• Cache-read input tokens: $0.0028 per 1M tokens
• Output tokens: $0.28 per 1M tokens

The cache-read rate is aimed at applications that reuse a stable system prompt, policy, product catalog or long reference context. Novita cautions that it is not an automatic discount on every request. Applications must follow the platform's current prompt-caching behavior, and users should verify how tokens are classified in their own account before basing architecture decisions on it.

The company also stresses that the listed rates are planning inputs rather than a billing guarantee, since real cost depends on prompt length, output length, caching behavior and retries. It advises rechecking the live model page before production budgeting.

Where Novita sees it fitting, and where not

Novita positions the model for high-frequency calls and large-scale professional workflows where cost matters alongside long-context and multimodal understanding. Suggested workloads include document-heavy agents, media-aware support flows, batch extraction and tool-using applications that need structured text responses.

The company is open about limits. A team handling short, text-only requests may prefer a smaller or more specialized endpoint. Teams that need a published benchmark result for a specific domain, or a response format not yet tested in their environment, may look elsewhere. Novita notes that its catalog provides capability and pricing metadata, not a workload-specific quality guarantee. It also says the 1M context figure is a maximum capacity, not a recommendation to fill every request, because very large prompts can raise cost and make failures harder to diagnose.

How Novita suggests testing it

The post recommends a small, deliberate first evaluation:

• Pick one measurable task with expected outputs, such as classification, extraction or a support workflow.
• Set a practical output cap, even though the model can return up to 128K tokens.
• Test modalities separately, starting with a text baseline before adding images, audio or video.
• Validate structured responses in the application and define a retry or repair path.
• Repeat representative requests with stable context and compare usage records to measure cache behavior.

For OpenAI-compatible integrations, developers set the model field to the exact ID and consult Novita's Create chat completion API reference for authentication and request details.

What it means for developers

For developers already on Novita AI, MiMo V2.6 Flash adds a low-priced option that combines long context, mixed media input and tool use in one serverless model. Novita points to Ling-3.0 Tiny and Qwen3.8 Flash as other models on its platform to compare against. The listing does not include benchmark results, so any decision to route production traffic to the model depends on teams running side-by-side tests with their own prompts, languages and media samples.