The Minimax M2.1 large language model is a sophisticated AI model primarily designed for coding and agentic workflows, building upon the foundation of its predecessor, M2 . This section details its technical specifications, architectural design, and key advancements.
Minimax M2.1 employs a sparse Mixture-of-Experts (MoE) transformer architecture . A key innovation is its efficient activation strategy, which allows for a high sparsity ratio . While the model comprises a total of 230 billion parameters, only 10 billion parameters are actively utilized during inference for each token . This design choice prioritizes inference throughput and local deployment on accessible hardware, such as consumer H100s or dual RTX 4090 setups, while maintaining a substantial "knowledge reservoir" 1. The architecture for M2, which M2.1 builds upon, includes a Multi-Head Attention (MHA) mechanism, characterizing it as a "full attention model" 2.
M2.1 supports a context window of 200,000 tokens , which is an increase from M2's 128,000 tokens . It also utilizes FP8 native quantization to optimize memory bandwidth usage against precision loss, demonstrating "computational pragmatism" in its engineering 1. A significant advancement in M2.1 is its implementation of "Advanced Interleaved Thinking," which enhances its systematic problem-solving capacity 3. This includes the use of ... tags to wrap reasoning content, which must be preserved in conversation history for optimal performance 4. The model is designed for concise, high-efficiency responses, reducing verbosity compared to previous generations and leading to faster "feel" and near-instant response times in developer workflows 3. While M2 initially supported multimodal inputs including text, audio, images, and video 2, M2.1's primary input type is listed as Text .
The training methodology for M2.1, similar to M2, involved an undisclosed data collection and labeling process 4. The data modality for training primarily consists of text and code 4. The model was trained with a strong emphasis on coding, agentic workflows, and tool use capabilities 4. For M2.1, specific optimizations have been applied for Web3 protocols, enhancing its performance in blockchain and decentralized projects 3. The model also shows advanced capabilities in multilingual coding, beyond just Python, covering languages such as Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, and JavaScript . The design supports end-to-end developer workflows, including multi-file edits, code-run-fix loops, and long-horizon toolchains , and provides native support for integrating external tools like shell environments, web browsers, and Python interpreters .
The key technical specifications are summarized in the following table:
| Feature | Value | Source |
|---|---|---|
| Architecture | Mixture-of-Experts (MoE) Transformer, Sparse MoE | |
| Total Parameters | 230 Billion | |
| Active Parameters | 10 Billion per token | |
| Context Window | 200,000 tokens | |
| Quantization | FP8 native 1 | |
| Input Modality | Text | |
| Attention Structure | Multi-Head Attention (MHA) 2 |
The Minimax M2.1 large language model, released on December 23, 2025, represents a substantial enhancement over its predecessor, M2, with a particular emphasis on improving performance in complex real-world tasks, especially in coding across multiple programming languages and office automation . It aims to achieve a leading position in these specialized domains 5.
Minimax M2.1 continues the "Mini" model philosophy designed for "Max" coding and agentic workflows, built as a compact, fast, and cost-effective Mixture-of-Experts (MoE) model. It utilizes 10 billion active parameters from a total of 230 billion for efficient performance, maintaining a streamlined form factor for easier deployment and scaling .
Minimax M2.1 introduces several key advancements over the M2 model, focusing on practical application and efficiency:
Minimax M2.1 delivers a significant leap over M2 on core software engineering leaderboards and demonstrates comprehensive improvements in specialized domains like test case generation, code performance optimization, code review, and instruction following .
Minimax M2.1 vs. M2 Performance Overview
| Benchmark | MiniMax-M2.1 Score | MiniMax-M2 Score |
|---|---|---|
| SWE-bench Verified | 74.0 | 69.4 |
| Multi-SWE-bench | 49.4 | 36.2 |
| SWE-bench Multilingual | 72.5 | 56.5 |
| Terminal-bench 2.0 | 47.9 | 30.0 |
| SWE-bench Verified (Claude Code) | 74.0 | 69.4 |
| SWE-bench Verified (Droid) | 71.3 | 68.1 |
| SWE-bench Verified (mini-swe-agent) | 67.0 | 61.0 |
| SWT-bench | 69.3 | 32.8 |
| SWE-Perf | 3.1 | 1.4 |
| SWE-Review | 8.9 | 3.4 |
| OctoCodingbench | 26.1 | 13.3 |
| VIBE (Average) | 88.6 | 67.5 |
| VIBE-Web | 91.5 | 80.4 |
| VIBE-Simulation | 87.1 | 77.0 |
| VIBE-Android | 89.7 | 69.2 |
| VIBE-iOS | 88.0 | 39.5 |
| VIBE-Backend | 86.7 | 67.8 |
| Toolathlon | 43.5 | 16.7 |
| BrowseComp | 47.4 | 44.0 |
| BrowseComp (context management) | 62.0 | 56.9 |
| AA-Index | 64.0 | 61.0 |
| MMLU | 88 | 82 (MMLU-Pro) |
| Humanity's Last Exam (HLE) w/o tools | 22.0 | 12.5 (M2) |
Comparative Analysis with Leading LLMs Minimax M2.1 shows strong competitiveness against other leading models including those from Anthropic, Google, and OpenAI .
Multilingual Coding and Software Engineering: M2.1 outperforms Claude Sonnet 4.5 in multilingual scenarios and approaches Claude Opus 4.5 3.
VIBE Benchmark (Full-Stack Development): M2.1 achieves an outstanding average score of 88.6 on the Visual & Interactive Benchmark for Execution (VIBE), demonstrating robust full-stack development capabilities . It particularly excels in the VIBE-Web (91.5) and VIBE-Android (89.7) subsets 5. M2.1's average VIBE score (88.6) is higher than GLM 4.6 (72.9) and Gemini 3 Pro (82.4), and close to Claude Sonnet 4.5 (85.2), but slightly lower than Claude Opus 4.5 (90.7) 3.
Tool Use and Agents: M2.1's Toolathlon score of 43.5 matches Claude Opus 4.5 (43.5) and outperforms Claude Sonnet 4.5 (38.9), Gemini 3 Pro (36.4), DeepSeek V3.2 (35.2), Kimi K2 Thinking (17.6), and GLM 4.6 (18.8) 3. For BrowseComp, M2.1 (47.4) beats Claude Sonnet 4.5 (19.6), Claude Opus 4.5 (37.0), Gemini 3 Pro (37.8), Kimi K2 Thinking (41.5), and GLM 4.6 (45.1), though GPT-5.2 (thinking) leads at 65.8 3.
General Intelligence and Knowledge: M2.1 scored 88 on MMLU, demonstrating strong knowledge and reasoning capabilities 6. This MMLU score is described as consistently equivalent to or closely behind flagship frontier models 6. While strong, other models like Reactor Mk.1 (92%) and GPT-4o (88.7%) have higher reported scores on MMLU 7. The AA Intelligence score for M2.1 is 64, an improvement over M2's 61 8. Minimax M2's composite intelligence score was previously ranked first among open-source models globally across mathematics, science, instruction following, coding, and agentic tool use 8.
Minimax M2.1 demonstrates significant strengths, particularly in its specific advancements over M2:
While M2.1 shows considerable strength, some areas indicate potential for further development or where other models may hold an advantage:
Information regarding Minimax M2.1's performance is primarily drawn from:
In summary, MiniMax M2.1 demonstrates significant advancements over its predecessor, M2, particularly in multilingual coding, app development, and agentic capabilities. Its high efficiency and competitive performance across a range of benchmarks position it as a powerful and cost-effective model, especially for developer-centric, high-volume, and complex problem-solving applications, challenging or outperforming many contemporary LLMs in various domains .
The Minimax M2.1 large language model is primarily engineered for advanced coding and complex agentic workflows, building upon the capabilities of its predecessor, M2 . Its unique architectural design, which includes a sparse Mixture-of-Experts (MoE) transformer with only 10 billion active parameters during inference despite a total of 230 billion parameters, prioritizes efficiency and local deployment on accessible hardware, making it a cost-effective state-of-the-art model . This efficiency, combined with a large context window of 200,000 tokens and FP8 native quantization, positions M2.1 for a wide array of demanding applications .
Minimax M2.1 demonstrates significant advancements in software engineering, particularly excelling in multilingual coding capabilities. It supports a comprehensive range of programming languages beyond Python, including Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript, and JavaScript . Its proficiency is evident in benchmarks such as SWE-bench Multilingual, where it scored 72.5%, surpassing many contemporary models like Claude Sonnet 4.5 and closely approaching Claude Opus 4.5 . The model is adept at various software development tasks including test case generation, code performance optimization, and code review .
Key coding performance benchmarks include:
| Benchmark | MiniMax-M2.1 Score |
|---|---|
| SWE-bench Verified | 74.0% |
| Multi-SWE-bench | 49.4% |
| SWE-bench Multilingual | 72.5% |
| Terminal-bench 2.0 | 47.9% |
A core strength of M2.1 lies in its capacity for agentic workflows and systematic problem-solving, driven by its "Advanced Interleaved Thinking" implementation . This feature allows the model to handle complex administrative and office automation tasks effectively, as demonstrated in the Toolathlon benchmark where it achieved a score of 43.5, matching Claude Opus 4.5 . M2.1 supports end-to-end developer workflows, encompassing multi-file edits, code-run-fix loops, and long-horizon toolchains . It also offers native support for integrating external tools like shell environments, web browsers, and Python interpreters, enhancing its utility in diverse development environments . The model shows exceptional performance across various programming agents and IDE extensions, including Claude Code, Droid (Factory AI), Cline, Kilo Code, and Roo Code .
Minimax M2.1 significantly boosts native Android and iOS development capabilities, along with enhancing "web aesthetics" for improved UI/UX generation and scientific scenario simulations . It excels in "vibe-coding" for aesthetically pleasing and functional UI designs across web and Android environments, with particular strengths in one-shot generation for Godot game engines and C++ graphics tasks 1. The model's full-stack development proficiency is rigorously evaluated by the VIBE (Visual & Interactive Benchmark for Execution) benchmark, where M2.1 achieved an outstanding average score of 88.6% . It performed exceptionally well in specific VIBE subsets: 91.5% in VIBE-Web, 89.7% in VIBE-Android, 88.0% in VIBE-iOS, and 86.7% in VIBE-Backend 5. The VIBE benchmark utilizes an Agent-as-a-Verifier (AaaV) paradigm to assess interactive logic and visual aesthetics in real runtime environments, underscoring M2.1's practical applicability in these domains 5.
M2.1 includes specific optimizations for Web3 protocols, enhancing its performance and applicability in blockchain and decentralized projects . This specialized focus enables the model to address the unique challenges and requirements of developing within the Web3 ecosystem.
The model's design for concise, high-efficiency responses, reducing verbosity compared to previous generations, results in a faster "feel" and near-instant response times for developer workflows . This efficiency, supported by its low activation parameter count and flexible deployment across various inference frameworks, ensures lower latency, reduced cost, and higher throughput for both interactive and batched workloads, making it ideal for dynamic development environments .
The Minimax M2.1 model is designed for developers managing diverse development scenarios, providing enhanced code quality, extensive coding scenario coverage across multiple languages, smarter instruction following, clearer reasoning, and cost efficiency for agentic workflows . This section details its availability, access methods, commercial models, and developer resources.
M2.1 is currently available in Preview for early access until December 22nd. Users can gain access by signing up or logging into the MiniMax Platform, obtaining a GroupId, and completing an early access request form 10.
API Key Acquisition and Endpoints: Users can obtain API keys either from the Account/Coding Plan page for subscribed coding plans, or by creating a new secret key on the MiniMax Developer Platform, which is displayed only once and must be securely saved 11. The base URL for the MiniMax API varies by region: international users should use api.minimax.io, while users in China should use api.minimaxi.com 11.
Integration Methods: Minimax M2.1 can be integrated into various coding tools and Command Line Interfaces (CLIs), typically requiring API key configuration and setting base URLs. It is crucial to clear any conflicting environment variables related to other AI providers such as Anthropic or OpenAI 11.
| Category | Tool/Method | Configuration Details |
|---|---|---|
| Recommended Integrations | Claude Code | Install Claude Code, configure ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN (with MiniMax API key), set model to MiniMax-M2.1 11. |
| Claude Code Extension for VS Code | Install the extension, set claude-code.selectedModel to MiniMax-M2.1, configure relevant environment variables 11. | |
| Cursor | Override the OpenAI Base URL, paste the MiniMax API key into the OpenAI API Key field, add MiniMax-M2.1 as a custom model 11. | |
| Other Integrations | TRAE | Add model, select "OpenRouter or SiliconFlow" as provider, "other models" for type, "MiniMax M2.1" as model ID, and MiniMax API key 11. |
| Droid | Configure ~/.factory/config.json with model_display_name, model (MiniMax-M2.1), base_url, api_key, provider (anthropic), and max_tokens 11. | |
| OpenCode | Configure ~/.config/opencode/opencode.json with baseURL and apiKey under the minimax provider, or use opencode auth login 11. | |
| Not Recommended | Codex CLI | Requires configuring .codex/config.toml with model_providers.minimax and profiles.m2.1, and setting MINIMAX_API_KEY environment variable 11. |
| Grok CLI | Involves setting GROK_BASE_URL and MINIMAX_API_KEY environment variables, then launching with grok --model MiniMax-M2.1 11. | |
| Coming Soon | Cline, Kilo Code, Roo Code | These tools currently support MiniMax-M2, with M2.1 integration planned for the future 11. |
Minimax offers a diverse commercial model, including a "Pay as You Go" API, specialized "Coding Plans," and "MiniMax Agent" plans, alongside options for audio and video services .
API Pricing (Pay as You Go): The API pricing for M2 (applicable to M2.1) is designed for cost-effectiveness, estimated at 8% of Claude 4.5 Sonnet's cost 12.
Coding Plans (Powered by MiniMax M2): These subscription packages, available by November 10th, are tailored for AI-powered coding .
| Plan | Price (Monthly) | Price (Yearly) | Prompts (Per 5 hours) | Features |
|---|---|---|---|---|
| Starter | $10 / month | $100 / year | 100 prompts | For entry-level developers managing lightweight workloads (equivalent to Claude Code Max 5x) . |
| Plus | $20 / month | $200 / year | 300 prompts | For professional developers managing complex workloads (3x Starter usage) 13. |
| Pro | $20 / month | $200 / year | 300 prompts | For professional developers managing complex workloads (equivalent to Claude Code Max 20x) 12. |
| Max | $50 / month | $500 / year | 1000 prompts | For power developers managing high-volume workloads (10x Starter usage; equivalent to Claude Code Max 20x) . |
MiniMax Agent Plans: These plans aim to optimize the economics of complex task completion by enabling agents to autonomously execute multi-turn searches, programming, and Office tool integrations 12.
Other Modalities: Minimax also provides independently developed ranges of modalities including text, audio, video, image, and music, each with flexible pricing plans to suit different usage requirements, such as Audio Subscription and Video Packages .
Minimax offers comprehensive resources to support developers integrating and utilizing the M2.1 model.
Developer Documentation: The MiniMax API Docs website serves as a central hub, providing developer guides, API references, pricing information, coding plans, solutions, release notes, and FAQs . Specific guides cover Quick Start, Models, Rate Limits, Text Generation, M2.1 Tool Use & Interleaved Thinking, M2.1 for AI Coding Tools, and Building Agents with M2: Best Practices 11.
SDKs/API Compatibility: Minimax M2.1 models are compatible with both the Anthropic API and OpenAI API for text generation. However, using the Anthropic SDK with MiniMax models is recommended for optimal integration 11.
Open-Source Aspects: While the M2.1 model itself is not explicitly described as open-source, the open-source model weights of MiniMax-M2 have been widely adopted on Hugging Face, showcasing community engagement and independent deployment across various platforms 12.
Key Features for Developers: M2.1 emphasizes improved code quality (leading to more readable and maintainable code), broader coverage for various coding scenarios, smarter instruction following, cleaner reasoning, and enhanced cost efficiency. These attributes make it particularly valuable for developing and implementing sophisticated agentic workflows 10.