Xiaomi MiMo's efficiency-focused omnimodal model, open-sourced on September 22, 2026. The API model id is mimo-v2.6-flash. The official model card lists 309B total parameters, 15B active parameters, and a 1M-token context. Availability in Tabbit depends on your account's live model picker.
Use in Tabbit
Use in Tabbit MiMo-V2.6-Flash
MiMo-V2.6-Flash · Model overview
Check the evidence before choosing a workflow
A compact view of reviewed task guides, public evaluations, and evidence boundaries. Client access still depends on your current account.
Task guides7
Review sources5
Sources reviewed12
Editor picks8
Model access and permissions must be checked in the current Tabbit account.
For mimo-v2.6-flash, first enable the Web Search Plugin in MiMo Console, then call the web_search tool through OpenAI Chat Completions; when real-time information is needed, use force_search: true, and use max_keyword to control the number of concurrent keywords per round and potential call costs.
Prepare
task goal, input material, tool or step constraints, acceptance criteria
Runtime
MiMo API's OpenAI Chat Completions; the documentation says that other API protocols are not currently supported.
MiMo-V2.6-Flash Deep Thinking Configuration and Multi-turn Tool-calling Workflow
mimo-v2.6-flash supports toggling deep thinking with thinking.type, which is enabled by default. When it is enabled, do not customize temperature or top_p, and pass through the historical assistant messages' reasoning_content in full during multi-turn tool calls.
Prepare
model ID, reasoning or call parameters, API endpoint, task, acceptance criteria
Runtime
The OpenAI Chat Completions-compatible interface of the MiMo API; in the Python SDK, thinking is not a standard OpenAI parameter and must be placed in extra_body.
MiMo-V2.6-Flash Structured Output: JSON Mode Configuration and Validation Workflow
The official documentation lists mimo-v2.6-flash as a model that supports JSON mode. When calling it, set response_format={"type": "json_object"} and explicitly require the system or user message to return JSON only, with fields, hierarchy, and types fully defined. This mode guarantees only valid JSON syntax, not the business structure, so production environments should still validate against a JSON Schema.
Prepare
model ID, reasoning or call parameters, API endpoint, task, acceptance criteria
Runtime
MiMo API; the source uses the OpenAI Python SDK and Curl examples, with base_url set to https://api.xiaomimimo.com/v1 in the examples.
MiMo-V2.6-Flash Image Understanding Inputs and Multi-image Workflow
The official documentation lists mimo-v2.6-flash as a supported image-understanding model. Images can be provided through a public URL or Base64, and multiple images can be compared; the documentation does not provide a Flash-specific response, so the Pro example output, token usage, and results shown on the page cannot be extrapolated to Flash.
Prepare
task goal, input material, tool or step constraints, acceptance criteria
Runtime
The MiMo API's OpenAI Chat Completions API; the source also provides the image content format for the Anthropic Messages API. The examples use the Python SDK and Curl.
MiMo-V2.6-Flash Official Benchmarks: 30 RL Steps and Agent Results
The vendor reports that after 30 RL steps and approximately 750,000 cumulative trajectories, MiMo-V2.6-Flash improved from 48.8 to 65.7 on DeepSWE v1.1; in the official Agent benchmark table, Flash scored 95.1 on CyberGym, 87.6 on Terminal Bench 2.1, and 71.5 on MiMo Visual Coding. All of these are results disclosed on Xiaomi's page, not independent reproductions.
Evidence
Vendor report
Boundary
These results cannot be used to infer performance across all real-world businesses, different prompts, different toolchains, or different inference parameters; the page does not provide the sample size for each benchmark, complete prompts, decoding settings, random seeds, hardware configuration。
MiMo-V2.6-Flash-RL Hugging Face Official Benchmarks and Deployment Boundaries
The official model card defines XiaomiMiMo/MiMo-V2.6-Flash-RL as the efficiency-balanced checkpoint in the MiMo-V2.6 series and reports its results on code, general Agent, cybersecurity, and visual Agent benchmarks. However, the evaluation hardware, sample sizes, complete harnesses, prompts, and decoding settings have not been disclosed.
Evidence
Vendor report
Boundary
Accuracy, latency, throughput, cost, and production stability for unlisted tasks or under different harnesses or decoding parameters; cross-model comparisons in the table cannot replace independent testing under matched conditions.
MiMo-V2.6-Flash Official X Release Thread: Flash's Benchmark Positioning and Dual-Model Strategy
The official post positions MiMo-V2.6-Flash as an omnimodal model alongside Pro, emphasizing scaled reinforcement learning, coding/general Agent/cybersecurity/visual Agent capabilities, and open reproduction; the attached image gives Flash's individual benchmark scores but does not disclose the sample size, complete harness, hardware, prompts, or decoding parameters。
Evidence
Vendor report
Boundary
Real-world business success rates, latency, throughput, cost, stability, or performance under different toolchains or inference parameters; Pro-exclusive claims must not be attributed to Flash.
