๐ผVentureBeatโขStalecollected in 17m
Xiaomi Launches MiMo Code: Agentic Coding for Long Tasks
๐กNew open-source coding agent claims to beat Claude Code on long-horizon tasks using a novel memory architecture.
โก 30-Second TL;DR
What Changed
Outperforms Claude Code on 200+ step complex coding tasks.
Why It Matters
This tool addresses the 'amnesia' problem in AI coding agents, potentially setting a new standard for how persistent memory is handled in long-running autonomous development workflows.
What To Do Next
Clone the MiMo Code repository from GitHub and test its performance on your most complex, multi-file refactoring task to see if the memory architecture reduces context loss.
Who should care:Developers & AI Engineers
Key Points
- โขOutperforms Claude Code on 200+ step complex coding tasks.
- โขFeatures a four-layer cross-session memory system powered by SQLite FTS5.
- โขUses an independent 'checkpoint-writer' subagent to manage project state and blueprints.
- โขAvailable via MIT license on GitHub with support for macOS, Linux, and Windows.
๐ง Deep Insight
Web-grounded analysis with 18 cited sources.
๐ Enhanced Key Takeaways
- โขMiMo Code is built as a fork of the open-source OpenCode project, with Xiaomi integrating its own memory architecture, workflow modes, and model harness to enhance its capabilities.
- โขThe tool provides free, limited-time access to Xiaomi's MiMo-V2.5 multimodal flagship model, which features a 1-million-token context window and is released under an MIT license.
- โขMiMo Code incorporates self-improvement mechanisms, including a "/dream" command that periodically (every seven days) reviews and compresses historical sessions into long-term memory, and a "distill" function that converts repeated workflows into reusable skills or commands.
- โขIt offers broad compatibility, supporting integration with various third-party AI services such as DeepSeek, Kimi, and GLM, which helps developers avoid vendor lock-in.
- โขThe assistant introduces a "Compose mode" that enables structured, specs-driven development workflows, orchestrating tasks from planning and design to coding, testing, and review based on a single goal prompt.
๐ Competitor Analysisโธ Show
Competitor Analysis: MiMo Code vs. Leading AI Coding Agents
| Feature/Metric | Xiaomi MiMo Code (with MiMo-V2.5-Pro) | Claude Code (with Opus 4.8) | OpenAI Codex CLI (with GPT-5.5) | Google Gemini CLI (with Gemini 3.1 Pro) |
|---|---|---|---|---|
| Availability | Open-source (MIT License) | Proprietary | Proprietary | Proprietary |
| Pricing | Free (limited-time free access to MiMo-V2.5) | Subscription ($20-200/month) | Free/Subscription/Usage-based | Free tier (60 req/min, 1,000/day) |
| Terminal-Native | Yes | Yes | Yes | Yes |
| Persistent Memory | Yes (4-layer cross-session, SQLite FTS5) | Yes (via plugins like claude-memory) | Varies by implementation | Varies by implementation |
| Long-Horizon Tasks | Designed for 200+ steps | Strong performance | Strong performance | Capable |
| SWE-Bench Pro | 62% (MiMo Code + MiMo-V2.5-Pro) | 69.2% (Opus 4.8) | 58.6% (GPT-5.5) | - |
| Terminal-Bench 2.1 | 73% (MiMo Code + MiMo-V2.5-Pro) | 78.9% (Opus 4.8) | 83.4% (GPT-5.5) | 70.7% (Gemini 3.1 Pro) |
| Underlying Model | MiMo-V2.5 / MiMo-V2.5-Pro | Claude Sonnet/Opus | GPT-5.5 | Gemini 3.1 Pro |
| Multi-Agent Support | Yes (build, plan, compose modes, subagents) | Yes (multi-agent workflows) | Yes (multi-agent worktrees) | Varies |
| Model Agnostic | Yes (supports DeepSeek, Kimi, GLM) | No (Claude-native) | No (OpenAI-native) | No (Gemini-native) |
Note: Benchmarks are self-reported by vendors unless otherwise specified and may be sensitive to configuration.
๐ ๏ธ Technical Deep Dive
- Memory Architecture: MiMo Code employs a four-layer cross-session memory system comprising Project Memory (stored in
MEMORY.mdfor persistent project knowledge, rules, and architectural decisions), Session Checkpoints (checkpoint.mdfor structured state snapshots managed by the 'checkpoint-writer' subagent), Scratch Notes (notes.mdfor temporary agent notes), and Per-Task Progress (tasks/<id>/progress.mdfor task-specific logs). - Persistent Storage: The memory system is powered by SQLite FTS5 (Full-Text Search), which enables efficient keyword-based retrieval and is favored over pure vector databases for its reviewability and local-first capabilities. Markdown files serve as the human-readable source of truth for project memory.
- Context Management: When resuming a session, the system rebuilds the agent's context by injecting information from these memory layers into a layered prompt. This prompt includes the task list, session checkpoint, recent user messages, project memory, global memory, notes, and an index of memory file paths, keeping the total injected content within approximately 65K tokens.
- Self-Improvement Mechanisms: MiMo Code features a
/dreamcommand that runs approximately every seven days to review old sessions, deduplicate memories, verify file paths, and compress information into an updated long-term memory repository. A/distillfunction is also available to identify and convert repeated work patterns into reusable skills or CLI commands. - Max Mode (Experimental): This feature enhances computation by sampling five independent plans in parallel for a given task and then using a judge model to select the most robust plan before execution. While it can improve performance by 10-20% on benchmarks like SWE-Bench Pro, it consumes roughly 4-5 times more tokens.
- Goal Verification: An independent judge model is used to verify whether the agent has successfully met the specified stopping conditions for a task, such as passing tests and committing code.
- Underlying Models: MiMo Code leverages Xiaomi's MiMo-V2.5 and MiMo-V2.5-Pro models. MiMo-V2.5 is a sparse Mixture-of-Experts (MoE) model with 310 billion total parameters (15 billion active per inference) and a 1-million-token context window, featuring a hybrid attention architecture. MiMo-V2.5-Pro is a larger 1.02-trillion-parameter MoE model with 42 billion active parameters.
- Dynamic Workflows: The orchestration logic within MiMo Code can be written as executable JavaScript, allowing for more reliable coordination of subagents, pipelines, retries, and parallel tasks.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
Increased adoption of open-source agentic coding tools.
Xiaomi's open-sourcing of MiMo Code with competitive performance and free access to its powerful MiMo-V2.5 model will likely drive more developers towards open-source alternatives, reducing reliance on proprietary solutions.
Accelerated innovation in long-horizon AI agent memory systems.
MiMo Code's novel cross-session memory architecture and explicit state management mechanisms, particularly the use of SQLite FTS5 and Markdown files, will likely inspire further research and development in persistent and reviewable memory solutions for AI agents.
Intensified competition among AI model providers to integrate with open-source agent frameworks.
MiMo Code's support for multiple third-party LLM providers (DeepSeek, Kimi, GLM) demonstrates a trend where agent frameworks become model-agnostic, pushing LLM developers to ensure compatibility and competitive performance within these ecosystems.
โณ Timeline
2025-04
Xiaomi releases MiMo-7B, its first open-source reasoning and coding model.
2025-11
Luo Fuli, a former DeepSeek researcher, joins Xiaomi to lead the MiMo division.
2025-12
Xiaomi releases MiMo-V2-Flash model.
2026-03
Xiaomi CEO Lei Jun announces a plan to invest US$8.7 billion in AI over the next three years.
2026-03-18
MiMo-V2-Pro (1.02T parameters) is launched, initially appearing anonymously as 'Hunter Alpha' on OpenRouter.
2026-04-22
Xiaomi officially launches MiMo-V2.5 and MiMo-V2.5-Pro models.
2026-06-10
Xiaomi open-sources MiMo Code V0.1.0, a terminal-native AI coding assistant.
๐ Sources (18)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: VentureBeat โ