1
3
Chat
Active
MiMo-V2.5 Pro
Whether you're running complex software engineering pipelines, long-horizon agentic tasks, or just pushing what a language model can reliably accomplish, this is where you start.
Xiaomi's most capable reasoning model to date, engineered for complex agentic workflows, long-horizon software engineering, and tasks that push the outer limits of what a language model can sustain on its own.
MiMo-V2.5-Pro isn't just a better version of its predecessor, it's a substantive re-think of what an AI model should be capable of when it has adequate time, context, and tools at its disposal. It succeeds MiMo-V2-Pro across every dimension that matters for real-world engineering work.
V2.5-Pro can maintain coherent, goal-directed behavior across thousands of tool calls within a single session without losing track of constraints, accumulated context, or intermediate state. That's a practical threshold most models can't approach.
The model demonstrates what the team calls "harness awareness", it actively maps its own environment, manages memory usage strategically, and shapes how context gets populated in service of the goal. It treats scaffolding as an extension of itself, not just a tool list.
With a 1M-token context window, V2.5-Pro doesn't just hold more information, it reasons over it coherently. Tasks spanning hours of real compute time remain structurally sound, with self-correcting behavior when intermediate results deviate from plan.
From repo-level understanding to project scaffolding, architecture planning, code review, and end-to-end builds, V2.5-Pro was evaluated against the Xiaomi MiMo Coding Bench, an in-house suite built to test exactly these real-world engineering workflows.
Synthetic benchmarks are useful signals, but the most honest test of a reasoning model is whether it can complete real work that would challenge a trained professional. Here's what MiMo-V2.5-Pro achieved when handed three demanding, open-ended tasks and left to run.
Task 01 · Systems ProgrammingThis task was drawn directly from Peking University's Compiler Principles course, where the reference project typically takes a CS major several weeks to complete. The model was asked to implement a full SysY language compiler in Rust: lexer, parser, abstract syntax tree, Koopa IR code generation, RISC-V assembly backend, and performance optimization.
Rather than approaching this through random trial and error, V2.5-Pro built the compiler layer by layer with deliberate architectural intent. It designed the full pipeline first, then systematically perfected each stage. The first compile alone passed 137 of 233 tests — a 59% cold-start pass rate that reflects how thoughtfully the architecture was planned before any test feedback was incorporated.
Around tool call 512, a refactoring pass introduced a regression in the RISC-V backend, temporarily dropping two tests. The model self-diagnosed the failure, traced it to root cause, corrected it, and continued — the kind of structured, self-correcting discipline that's essential for any task stretching across hundreds of incremental steps.
Task 02 · Desktop Application DevelopmentStarting from a few simple text prompts, V2.5-Pro autonomously built a fully functional desktop video editing application: multi-track timeline, clip trimming, cross-fade transitions, audio mixing, and a complete export pipeline. The output was not a stub or a mockup — it was a working application.
This task is notable not just for its scale but for what it required the model to manage internally: UI state, media pipeline architecture, inter-component data flow, feature creep avoidance, and progressive testing of individual components before integration. The final build also incorporated an AI voice-over track powered by MiMo-V2-TTS, demonstrating the model's ability to orchestrate multiple AI subsystems as part of a broader product.
Task 03 · Analog Circuit DesignThis graduate-level analog EDA task asked the model to design and optimize a complete Flipped-Voltage-Follower low-dropout regulator (FVF-LDO) from scratch in a real 180nm semiconductor process. The challenge: six competing metrics must land within specification simultaneously — phase margin, line regulation, load regulation, quiescent current, PSRR, and transient response. Trained analog designers typically spend several days on a project at this scope.
V2.5-Pro was wired into an ngspice simulation loop using Claude Code as the execution harness. Through closed-loop iteration — calling the simulator, reading waveforms, adjusting component sizing and bias voltages — the model converged on a design where every target metric was met. Four of the key metrics improved by roughly an order of magnitude compared to its own first attempt.
This task exemplifies the "harness awareness" behavior that defines V2.5-Pro's character: it understood the simulation environment as a feedback mechanism, used it deliberately, and built an engineering intuition for how parameter changes propagated through the circuit — rather than treating each iteration as an isolated guess.
To evaluate models fairly on real engineering workflows, Xiaomi developed the MiMo Coding Bench — an internal evaluation suite covering a broad spectrum of developer scenarios, from isolated code review to full-scale project construction within agentic frameworks like Claude Code.
The suite tests models across: repository understanding, project construction, code review, structured artifact generation, planning, software engineering, and more — all within the constraints and dynamics of real agentic execution flows, not isolated prompts.
MiMo-V2.5-Pro has substantially narrowed the gap to Claude Opus 4.6 on MiMo Coding Bench — the top-performing proprietary model in agentic coding evaluations. This makes V2.5-Pro a credible alternative for developers integrating into scaffolds like Claude Code, OpenCode, and Kilo Code at meaningfully lower per-token cost.
MiMo-V2.5-Pro isn't just a better version of its predecessor, it's a substantive re-think of what an AI model should be capable of when it has adequate time, context, and tools at its disposal. It succeeds MiMo-V2-Pro across every dimension that matters for real-world engineering work.
V2.5-Pro can maintain coherent, goal-directed behavior across thousands of tool calls within a single session without losing track of constraints, accumulated context, or intermediate state. That's a practical threshold most models can't approach.
The model demonstrates what the team calls "harness awareness", it actively maps its own environment, manages memory usage strategically, and shapes how context gets populated in service of the goal. It treats scaffolding as an extension of itself, not just a tool list.
With a 1M-token context window, V2.5-Pro doesn't just hold more information, it reasons over it coherently. Tasks spanning hours of real compute time remain structurally sound, with self-correcting behavior when intermediate results deviate from plan.
From repo-level understanding to project scaffolding, architecture planning, code review, and end-to-end builds, V2.5-Pro was evaluated against the Xiaomi MiMo Coding Bench, an in-house suite built to test exactly these real-world engineering workflows.
Synthetic benchmarks are useful signals, but the most honest test of a reasoning model is whether it can complete real work that would challenge a trained professional. Here's what MiMo-V2.5-Pro achieved when handed three demanding, open-ended tasks and left to run.
Task 01 · Systems ProgrammingThis task was drawn directly from Peking University's Compiler Principles course, where the reference project typically takes a CS major several weeks to complete. The model was asked to implement a full SysY language compiler in Rust: lexer, parser, abstract syntax tree, Koopa IR code generation, RISC-V assembly backend, and performance optimization.
Rather than approaching this through random trial and error, V2.5-Pro built the compiler layer by layer with deliberate architectural intent. It designed the full pipeline first, then systematically perfected each stage. The first compile alone passed 137 of 233 tests — a 59% cold-start pass rate that reflects how thoughtfully the architecture was planned before any test feedback was incorporated.
Around tool call 512, a refactoring pass introduced a regression in the RISC-V backend, temporarily dropping two tests. The model self-diagnosed the failure, traced it to root cause, corrected it, and continued — the kind of structured, self-correcting discipline that's essential for any task stretching across hundreds of incremental steps.
Task 02 · Desktop Application DevelopmentStarting from a few simple text prompts, V2.5-Pro autonomously built a fully functional desktop video editing application: multi-track timeline, clip trimming, cross-fade transitions, audio mixing, and a complete export pipeline. The output was not a stub or a mockup — it was a working application.
This task is notable not just for its scale but for what it required the model to manage internally: UI state, media pipeline architecture, inter-component data flow, feature creep avoidance, and progressive testing of individual components before integration. The final build also incorporated an AI voice-over track powered by MiMo-V2-TTS, demonstrating the model's ability to orchestrate multiple AI subsystems as part of a broader product.
Task 03 · Analog Circuit DesignThis graduate-level analog EDA task asked the model to design and optimize a complete Flipped-Voltage-Follower low-dropout regulator (FVF-LDO) from scratch in a real 180nm semiconductor process. The challenge: six competing metrics must land within specification simultaneously — phase margin, line regulation, load regulation, quiescent current, PSRR, and transient response. Trained analog designers typically spend several days on a project at this scope.
V2.5-Pro was wired into an ngspice simulation loop using Claude Code as the execution harness. Through closed-loop iteration — calling the simulator, reading waveforms, adjusting component sizing and bias voltages — the model converged on a design where every target metric was met. Four of the key metrics improved by roughly an order of magnitude compared to its own first attempt.
This task exemplifies the "harness awareness" behavior that defines V2.5-Pro's character: it understood the simulation environment as a feedback mechanism, used it deliberately, and built an engineering intuition for how parameter changes propagated through the circuit — rather than treating each iteration as an isolated guess.
To evaluate models fairly on real engineering workflows, Xiaomi developed the MiMo Coding Bench — an internal evaluation suite covering a broad spectrum of developer scenarios, from isolated code review to full-scale project construction within agentic frameworks like Claude Code.
The suite tests models across: repository understanding, project construction, code review, structured artifact generation, planning, software engineering, and more — all within the constraints and dynamics of real agentic execution flows, not isolated prompts.
MiMo-V2.5-Pro has substantially narrowed the gap to Claude Opus 4.6 on MiMo Coding Bench — the top-performing proprietary model in agentic coding evaluations. This makes V2.5-Pro a credible alternative for developers integrating into scaffolds like Claude Code, OpenCode, and Kilo Code at meaningfully lower per-token cost.
-p-130x130q80-p-130x130q80.png)
-p-130x130q80.png)
.png){kind=small}
