Xiaomi's MiMo Code scales coding agents to long-horizon tasks
Xiaomi's MiMo team has open-sourced MiMo Code, a terminal-based coding agent built on OpenCode and designed to maintain decision quality and state continuity across dozens or even hundreds of execution steps.
Score breakdown
MiMo Code's parallel sampling and selection approach demonstrates a concrete, measurable tradeoff — a 10–20% SWE-Bench Pro gain at 4–5× token cost — for improving reliability in long-horizon agentic coding runs where compounding step errors and context degradation are otherwise unmitigated.
- 01MiMo Code is a terminal-based coding agent built by Xiaomi's MiMo team on top of OpenCode, open-sourced under the MIT license.
- 02It is designed for long-horizon automated programming tasks spanning dozens or even hundreds of execution steps.
- 03The design is organized around three themes: computation, memory, and evolution.
Xiaomi's MiMo team has open-sourced MiMo Code, a terminal-based coding agent built on OpenCode and released under the MIT license. The project targets long-horizon automated programming tasks — those spanning dozens or even hundreds of execution steps — and is organized around three themes: computation, memory, and evolution, each corresponding to a distinct bottleneck the team identified at different time scales.
The core design motivation stems from two problems that emerge as task length grows.
The core design motivation stems from two problems that emerge as task length grows. First, context windows eventually fill with tool outputs, code snippets, and error logs, forcing compression or discard of history. The post argues that simple summarization creates a recurrent-model-style dilemma — it cannot look back on demand — and that what is needed instead is an explicit storage-and-retrieval mechanism. Second, even with a large enough context window, a model's instruction-following ability degrades as input length grows, with useful constraints diluted by accumulated tool output.
On the computation theme, MiMo Code introduces "Max Mode," which generates N candidate solutions in parallel at each turn (N defaults to 5, at temperature 1), then uses the same model as a judge to select the best reasoning and action plan before any execution occurs. On SWE-Bench Pro, Max Mode improves performance by 10–20% compared with single sampling, at a cost of roughly 4–5× token consumption. Max Mode is currently experimental and must be enabled manually. A second mechanism called "Goal" addresses premature task termination — a failure mode where the agent incorrectly declares completion after observing prior progress — which is described as especially dangerous in automated execution where no human is available to provide corrective feedback. The source text is truncated before the full details of Goal and the memory and evolution themes are described.
Key facts
- 01MiMo Code is a terminal-based coding agent built by Xiaomi's MiMo team on top of OpenCode, open-sourced under the MIT license.
- 02It is designed for long-horizon automated programming tasks spanning dozens or even hundreds of execution steps.
- 03The design is organized around three themes: computation, memory, and evolution.
- 04Max Mode generates N candidate solutions in parallel per turn (N defaults to 5, at temperature 1) and uses the same model as a judge to select the best plan before execution.
- 05Max Mode improves performance by 10–20% on SWE-Bench Pro compared with single sampling, at roughly 4–5× the token cost.
- 06Max Mode is currently experimental and must be enabled manually through configuration.
- 07A 'Goal' mechanism addresses premature task termination, a failure mode where the agent incorrectly declares completion during long automated runs.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 14, 2026 · 09:08 UTC. How this works →