MiMo V2.6: Pro, Flash and the Cost of a Finished Task

Use mimo-v2.6-pro for the flagship API and mimo-v2.6-flash for the lower-cost route. Both list multimodal understanding, text generation, thinking, function calls, structured output and web search, with a 1M context and 128K maximum output. The separately named mimo-v2.6-pro-ultraspeed is a premium serving route; its API tariff is ten times ordinary Pro’s.

Sign into Xiaomi’s console, enable the intended billing route and create its credential. The ordinary API uses the OpenAI-compatible base https://api.xiaomimimo.com/v1, with an Anthropic-compatible endpoint also available. Token Plan and Batch have dedicated endpoints and credentials supplied by their console flows. A desktop membership, a Token Plan pool and a pay-as-you-go balance are separate purchases.

The request below uses the ordinary API and an explicit output ceiling. Inspect returned content and usage before scaling it into an agent loop. An agent client supplies the repository, tools and execution permissions; the model request alone does not run tests.

These are overseas pay-as-you-go USD rates per million tokens, checked October 14. Cache-hit input is billed separately from misses; cache writes are temporarily free. Domestic billing uses CNY. Web search is additional: $5 per 1,000 overseas calls or ¥16 per 1,000 domestic calls.

For an assumed 20,000 uncached input tokens and 2,000 output tokens, Flash costs $0.00336, Pro $0.01044 and UltraSpeed $0.1044. If 18,000 input tokens are actual cache hits, Flash becomes $0.0008904 and Pro $0.0026748. Sum all reasoning/output, tool charges and retries before dividing by accepted results.

Batch halves the listed Pro and Flash token rates; UltraSpeed has no Batch route. The uncached example becomes $0.00168 for Flash and $0.00522 for Pro. Use Batch’s separate submission/status/result workflow for asynchronous jobs, keeping request identifiers with the results. Its lower price is a different delivery path from interactive streaming.

The individual monthly plans list Lite at $6 for 4.1B credits, Standard $16 for 11B, Pro $50 for 38B and Max $100 for 82B. Credits are converted by model and token type: Pro consumes 2.5 per cached input token, 300 per uncached input and 600 per output; Flash uses 2, 100 and 200.

The same assumed uncached 20K-input/2K-output request consumes 2.4M credits on Flash or 7.2M on Pro. If the entire Lite quota served only that request shape, it would cover 1,708 complete Flash requests or 569 Pro requests. This calculation excludes other usage and cache effects; 4.1B credits is not 4.1B tokens.

When the pool is exhausted, service pauses rather than silently drawing from the ordinary API balance. Upgrade the plan or deliberately switch to the metered API to continue. The current plan table lists Pro and Flash, so do not assume UltraSpeed is included. Choose a subscription when its usable workload and limits fit the month, not by dividing its fee by an advertised token count.

The October 14 change is an old-ID routing transition, not a new V2.6 release. mimo-v2.5-pro maps to mimo-v2.6-pro and mimo-v2.5 maps to mimo-v2.6-flash. Old IDs stop working on October 21 at 02:00 UTC. The official transition table says 02:00 UTC on October 14, while its note says 10:00 UTC; use the new IDs directly rather than depending on that conflicting hour.

For a migrated Flash request without max_completion_tokens, the documented default rises from 32,768 to 131,072. Set an explicit ceiling for the task and compare the resulting output usage. In thinking mode, the new models use temperature 1.0 and top_p 0.95; custom values are not supported there.

The official MIT-licensed RL checkpoints are XiaomiMiMo/MiMo-V2.6-Pro-RL and XiaomiMiMo/MiMo-V2.6-Flash-RL. Pro’s card describes a 1.02T-parameter MoE with 42B activated per token; both cards describe text, image, video and audio support with a 1M context. The collection also contains MOPD and a distinct 9B distilled checkpoint. Choose the exact artifact and its recipe rather than substituting a smaller derivative into a Pro API comparison.

Pro’s published vLLM example uses tensor parallelism across eight devices; its SGLang example is a multi-node configuration. These are deployment recipes, not a universal minimum device count. Active parameters reduce work per token but do not remove the need to place the full selected weights, encoders and runtime buffers somewhere.

For occasional use, the API avoids maintaining that serving stack. For self-hosting, budget the complete GPU session, storage and operations and measure accepted tasks per hour. This page has no hardware calculator preset because a trustworthy device/format profile has not been established.

Xiaomi’s model-card comparison reports DeepSWE v1.1 at 71.9 for Pro and 67.9 for Flash, but Terminal Bench 4.0 at 34.9 and 28.8. They measure different workloads; these vendor-reported results do not establish a same-cost winner on your repository. UltraSpeed’s advertised acceleration concerns its serving mode, not a guarantee that a finished agent task takes one twentieth of the time.

Keep the tools, inputs and acceptance checks fixed when comparing Flash with Pro or a hosted alternative. A successful extraction or accepted patch is the unit to budget; prompt length, reasoning and retries determine whether a lower token rate remains cheaper. For rigid OCR or timestamped parsing, Youtu’s task-specific JSON workflow is a different option from asking a general agent to interpret a file.