Xiaomi MiMo V2.6-Pro

Overview

Xiaomi MiMo V2.6-Pro is the flagship model in the MiMo V2.6 series, released and open-sourced on September 22, 2026. It retains MiMo V2.5-Pro's 1M context and 128K maximum completion budget (131,072 tokens including reasoning tokens, according to the Chat Completions API reference), but moves the Pro tier to a native multimodal architecture that can work with text, images, video, and audio. Xiaomi targets long-horizon coding, research, cybersecurity, design, and computer-operation agents rather than short single-turn prompts.

The upgrade case rests on scaled mixed-task reinforcement learning, stronger official agent benchmarks, and immediate access to MIT-licensed weights. API users also get the mimo-v2.6-pro model through OpenAI- and Anthropic-compatible endpoints, while local teams can deploy the official checkpoint with SGLang or vLLM. Xiaomi's benchmark and cost claims remain vendor-reported, so production teams should compare both quality and total trajectory cost on their own workloads.

What's New

Native multimodal capability moves into the Pro model

MiMo V2.5 split its positioning between a long-horizon Pro model and a less expensive native multimodal model. V2.6-Pro combines those priorities: Xiaomi and its official model card describe one model that accepts text, image, video, and audio context while handling coding, visual creation, research, office work, and security tasks.

The context window remains 1 million tokens, so this is not a capacity increase. The practical change is that teams can test a single Pro checkpoint for long tool traces and mixed-media evidence instead of routing perception and agent reasoning through separate MiMo models.

Reinforcement learning scales across tasks and agent harnesses

Xiaomi trained Pro and Flash in one mixed environment spanning coding, general agents, visual tasks, and cybersecurity. The official report says each update used 1,568 prompts with 16 rollouts per prompt, while multi-harness training exposed the models to different system prompts, tools, and context-management strategies.

During the six-day live run, Xiaomi reports that V2.6-Pro completed 30 RL steps, cost about $2.62 million to train, improved its average training-task pass rate by 12% relative, and raised an out-of-sample DeepSWE v1.1 result from 58.4 to 72.6. The announcement reports 72.6 for the live run, while the model-card table reports 71.9; those sources do not explain the discrepancy. The live-run starting score of 58.4 is also distinct from the table's V2.5-Pro baseline of 19.0.

Groupwise grading favors better successful trajectories

The V2.6 technical materials describe Groupwise Reward Synthesis and Groupwise Advantage Redistribution. Instead of treating every passing solution as equally good, an agentic grader builds task-specific rubrics, compares successful trajectories, and shifts learning weight toward higher-quality paths.

For users, the intended benefit is more reliable long-horizon execution with fewer unnecessary steps. It is still a training design claim, not a guarantee that every agent framework will use fewer tokens or finish more tasks.

Open weights, report, environments, and RL tooling arrive together

Xiaomi released the V2.6-Pro-RL and V2.6-Flash-RL weights, a distilled Qwen 9B model, the technical report, more than 7,000 RL task environments, and supporting training resources. The official V2.6-Pro-RL repository is marked with an MIT license.

Xiaomi had already released V2.5-Pro weights under MIT before V2.6, as documented in its V2.5 open-source announcement. The distinction is release-day availability of the new V2.6 RL checkpoint and its accompanying research resources, not the first availability of open Pro weights. Researchers can inspect and self-host the checkpoint instead of relying only on Xiaomi's hosted API.

UltraSpeed and new first-party clients broaden access

The standard mimo-v2.6-pro model is available through Xiaomi's Open Platform, AI Studio, MiMo Code, MiMo Desktop, and OpenRouter. Xiaomi also lists mimo-v2.6-pro-ultraspeed, a hosted mode advertised at up to 20 times the standard inference speed for latency-sensitive interaction.

MiMo Desktop and a membership plan launched alongside the models. The announcement says the invitation-only beta ends one week after release, while users who already qualified can continue after switching model names.

Performance Benchmarks

The official open-weight model card reports the following results. Scores depend on Xiaomi's stated harnesses and settings and should be treated as vendor evidence rather than independent production forecasts.

The table supports two different decisions. V2.6-Pro is the higher-scoring MiMo option on most listed agent tasks, but V2.6-Flash remains close on several rows and can be the more economical choice. Against closed models, Pro scores above both listed closed models on AutomationBench and Terminal Bench 2.1, but below both on DeepSWE. On Toolathlon-Verified, its 76.9 is below Claude Opus 5's 80.6 but above GPT-5.6 Sol's 74.9. Xiaomi's full model card also reports lower ExploitGym and ExploitBench scores than both.

Compared With MiMo V2.5-Pro

Migration & Compatibility

Hosted API users should change the model identifier to mimo-v2.6-pro. Xiaomi documents both OpenAI- and Anthropic-compatible interfaces, so existing MiMo users staying on the same protocol and billing channel can generally retain their SDK, base URL, and credentials while changing the model name and testing compatibility. Moving from another provider or between pay-as-you-go and Token Plan may also require different connection settings. The one-week early-access extension concerns MiMo Desktop and should not be confused with the API retirement schedule. Xiaomi's deprecation notice schedules retirement of mimo-v2.5-pro and mimo-v2.5 for October 21, 2026 at 10:00 Beijing time, without automatic replacement; requests using those IDs after retirement will return errors. Test and complete the hosted API migration before that deadline.

The official model page lists deep thinking, tool calling, streaming, web search, structured output, and context caching. Its visible specification labels the input modality as text even though the same page, release announcement, and model card describe native image, video, and audio understanding. Developers should confirm the current request schema and endpoint support for each non-text modality before migrating a production multimodal workflow.

Self-hosting is a substantially different operational choice. Xiaomi's SGLang example uses tensor parallelism across 16 devices, data parallelism across two groups, and expert parallelism across 16 devices; its vLLM example uses tensor parallelism across eight devices. The model card does not provide a universal minimum VRAM figure, so teams should size hardware against the selected weight format, context target, concurrency, and multimodal workload.

Pricing & Access

The official MiMo V2.6-Pro model page currently lists the following pay-as-you-go rates per 1 million tokens:

The page also lists a 100 RPM and 10-million-TPM limit. Token Plan subscriptions cover V2.6-Pro, V2.6-Flash, and the V2.5 family, but subscription economics depend on the selected plan and usage. UltraSpeed is available through MiMo Desktop and the Open Platform; Xiaomi advertises up to 20x output speed. The announcement's pricing graphic and the official API pricing page list UltraSpeed at $0.036 cached input, $4.35 cache-miss input, and $8.70 output per 1M tokens (¥0.25, ¥30, and ¥60 respectively). These are 10x the standard Pro token rates.

Hosted access is available through Xiaomi's API, AI Studio, MiMo Code, MiMo Desktop, MiMo Claw, and OpenRouter. For local deployment, the official Hugging Face checkpoint provides the weights and examples for SGLang and vLLM.

Who Should Upgrade / Who Should Wait

Upgrade if:

• You use V2.5-Pro for long coding or research agents and want Xiaomi's stronger official agent scores without giving up 1M context.
• You need one Pro model to reason over text, images, video, and audio instead of routing multimodal work through a separate V2.5 model.
• You specifically need the V2.6-Pro-RL checkpoint's capabilities and can operate its distributed inference infrastructure; MIT licensing alone is not a reason to upgrade from the already-open V2.5-Pro.
• You use OpenAI- or Anthropic-compatible clients and can run a controlled model-name and endpoint migration.

Wait if:

• You self-host V2.5 and it already meets your requirements. Users of Xiaomi's hosted mimo-v2.5-pro and mimo-v2.5 APIs must still migrate before their scheduled October 21, 2026 retirement, even when satisfied with the older models.
• You require independent benchmark reproduction before adopting Xiaomi's vendor-published agent and security results.
• Your self-hosting environment cannot support the model's distributed inference examples, and hosted deployment is not acceptable.
• Your production workflow depends on image, audio, or video API schemas that Xiaomi's public model page does not yet describe consistently.

Known Limitations

• The 1M context window and 128K maximum output do not increase over MiMo V2.5-Pro.
• Benchmark results come from Xiaomi's release materials and may use harnesses or settings that differ from production.
• The public model page's text-only input label conflicts with its native multimodal capability description, so endpoint-level compatibility needs testing.
• Self-hosting the 1.02T-parameter checkpoint requires distributed infrastructure; the official card does not promise a single minimum hardware profile.
• UltraSpeed has published token prices at 10x standard Pro rates and customized rate limits, rather than the standard model's listed 100 RPM / 10M TPM. Its advertised up-to-20x gain concerns output speed, not a guaranteed reduction in end-to-end task latency.

FAQ

Sources

• Xiaomi MiMo V2.6 release announcement — release date, training process, access, UltraSpeed, open-source scope, and vendor performance claims.
• MiMo V2.6-Pro model page — current specifications, token prices, rate limits, capabilities, and API example.
• MiMo V2.6-Pro-RL model card — MIT-licensed weights, architecture, benchmark table, and deployment guidance.