개요
MiniMax M2.5 is a high-performance Mixture-of-Experts model with 230 billion total parameters and 10 billion active, released on February 12, 2026. It achieves state-of-the-art results in coding, agentic tool use, and office productivity, scoring 80.2% on SWE-Bench Verified and 76.3% on BrowseComp, while being dramatically cost-effective at $0.15 per million input tokens for the standard variant. Trained via large-scale reinforcement learning across 200,000+ real-world environments using MiniMax's Forge framework, M2.5 excels in task decomposition and efficiency, completing agentic tasks 37% faster than its predecessor M2.1.
Positioned as a frontier-competitive open-weight model, M2.5 targets developers and enterprises needing high-volume, cost-sensitive coding and automation workflows. Its MIT license allows for self-hosting and customization, challenging closed-source models by offering near-matching performance at a fraction of the cost. However, it has notable limitations including a high hallucination rate and lack of multimodal capabilities, making it a specialist for coding and agentic tasks rather than a general-purpose model.
벤치마크 및 성능
MiniMax M2.5 delivers strong benchmark performance across coding, agentic, and reasoning tasks, as shown in the table below (scores from official evaluations and independent reviews). All benchmarks are from the MiniMax official announcement unless noted.
| Benchmark | MiniMax M2.5 | Claude Opus 4.6 | GPT-5.2 (thinking) | Gemini 3 Pro | Notes |
|----------------------------|--------------|-----------------|-------------------|--------------|-------|
| SWE-Bench Verified | 80.2% | 80.8% | 80.0% | 78% | Highest open-weight score, close to frontier |
| Multi-SWE-Bench | 51.3% | 50.3% | 49.1% | 42.7% | First place, multilingual coding tasks |
| BrowseComp (with context) | 76.3% | 84.0% | 65.8% | 73.2% | Leading open-weight for web search tasks |
| GPQA-D | 85.2% | 90.0% | 90.0% | 91.0% | Graduate-level scientific reasoning |
| AIME25 | 86.3 | 95.6 | 98.0 | 96.0 | Math reasoning, lags behind frontier |
| IFBench | 70.0% | 53.0% | 75.0% | 70.0% | Instruction following |
| BFCL Multi-Turn | 76.8% | 63.3% | - | - | Tool calling, outperforms Claude |
| Intelligence Index (Artificial Analysis) | 42 | 53 | 57 | 57 | Overall capability ranking |
Additional notes: M2.5-Lightning variant offers 100 tokens/sec throughput, completing SWE-Bench tasks 37% faster than M2.1. On agentic tasks, it uses 20% fewer rounds than predecessors, improving efficiency. The model's hallucination rate is high at 88% per Artificial Analysis, which impacts factual reliability.
상세 비교
**MiniMax M2.5 vs. Claude Opus 4.6:**
- **Pricing:** M2.5 is drastically cheaper at $0.15/$1.20 per 1M tokens vs. $5.00/$25.00 for Opus 4.6, making it 10-20x more cost-effective for output-heavy workloads.
- **Context Window:** Both offer 200K tokens, but M2.5 supports a 1M token context in its description (though practical limits are 200K per evaluations).
- **Strengths:** M2.5 excels in cost efficiency, speed (100 tokens/sec vs. ~60), and open-weight flexibility. It leads on Multi-SWE-Bench and BFCL Multi-Turn tool calling. Opus 4.6 has higher scores in BrowseComp (84.0% vs. 76.3%) and creative reasoning.
- **Weaknesses:** M2.5 has a high hallucination rate and lags in math reasoning. Opus 4.6 is proprietary with better factual reliability and multimodal support.
**MiniMax M2.5 vs. GPT-5.2:**
- **Pricing:** M2.5 is cheaper than GPT-5.2, which costs ~$1.25/$10.00 per 1M tokens.
- **Context Window:** M2.5's 200K tokens is smaller than GPT-5.2's 400K tokens.
- **Strengths:** M2.5 has comparable SWE-Bench scores (80.2% vs. 80.0%) and superior tool use (BFCL 76.8% vs. lower for GPT-5.2). It's open-weight and faster.
- **Weaknesses:** GPT-5.2 outperforms in math (AIME 25: 98.0 vs. 86.3) and multimodal tasks. M2.5's hallucination rate is higher.
**MiniMax M2.5 vs. Gemini 3 Pro:**
- **Pricing:** M2.5 is cheaper at $0.15/$1.20 vs. $2.00/$8.00 for Gemini 3 Pro.
- **Context Window:** Gemini 3 Pro supports up to 1M tokens, while M2.5 is 200K.
- **Strengths:** M2.5 leads in Multi-SWE-Bench (51.3% vs. 42.7%) and is more cost-effective. Gemini 3 Pro has higher GPQA (91.0% vs. 85.2%) and better factual reliability.
- **Weaknesses:** Gemini 3 Pro is proprietary with multimodal capabilities; M2.5 is text-only and has higher hallucination.
커뮤니티 평가
Developer and researcher reactions to MiniMax M2.5 highlight its cost-effectiveness and coding prowess, but also raise concerns about reliability and controversies. On platforms like Hacker News and GitHub, developers praise its SWE-Bench scores and pricing, noting it as a game-changer for budget-constrained teams. The OpenHands index ranked it 4th overall for long-running tasks, with feedback indicating it 'delivers a flawless closed-loop system' on multi-file development.
However, skepticism exists around the high hallucination rate (88%) reported by Artificial Analysis, which impacts trust in factual tasks. The distillation controversy, where Anthropic accused MiniMax of using synthetic Claude interactions for training, has sparked debate on ethical sourcing and benchmark integrity. Community adoption is high, with M2.5 becoming the most-used model on OpenRouter within a week of release, handling 2.45 trillion tokens. MiniMax's internal use, where 80% of new code is AI-generated by M2.5, underscores its practical utility.
활용 사례
**1. Software Engineering and Code Generation:** M2.5 is ideal for automating GitHub PRs, bug fixes, and full-stack app development. For example, building a React app with Node.js backend results in a complete spec-first plan and functional code in ~22 minutes, leveraging its 80.2% SWE-Bench score. Choose M2.5 over alternatives when cost is critical and tasks involve multi-language codebases (10+ languages supported).
**2. Agentic Pipelines and Tool Use:** With strong BFCL Multi-Turn (76.8%) and BrowseComp scores, M2.5 excels in multi-step agent workflows like automated research, web scraping, or API orchestration. It uses 20% fewer rounds than predecessors, making it efficient for high-volume pipelines. Use it over Claude Opus 4.6 when latency and cost matter more than absolute accuracy.
**3. Office Productivity and Document Automation:** Trained with industry experts, M2.5 generates deliverable outputs for Word, Excel, and PowerPoint, achieving a 59% win rate on GDPval-MM. For financial modeling or report generation, it follows structured SOPs to produce professional results. Choose it over GPT-5.2 when needing cost-effective, domain-specific office automation.
**4. Scientific Research and Q&A:** With an 84.8% GPQA-D score, M2.5 can handle graduate-level scientific queries and literature synthesis. It's suitable for building research assistants or tutoring tools where budget constraints apply. Opt for M2.5 over models like Gemini 3 Pro for lower costs, but use proprietary models for critical factual retrieval due to hallucination risks.
최신 뉴스
Recent developments for MiniMax M2.5 include:
- **API Pricing and Plans:** M2.5 is available in two variants: Standard ($0.15/$1.20 per 1M tokens at 50 tokens/sec) and Lightning ($0.30/$2.40 per 1M tokens at 100 tokens/sec), with caching support. A coding plan is offered via the MiniMax platform.
- **Forge RL Framework:** MiniMax detailed its agent-native reinforcement learning framework Forge, which enabled rapid training across 200,000+ environments, contributing to M2.5's efficiency improvements.
- **Distillation Controversy:** In February 2026, Anthropic accused MiniMax of using over 13 million synthetic Claude interactions via 24,000 fraudulent accounts to train M2.5, sparking industry debate on ethical AI training.
- **Rapid Improvement:** MiniMax released M2, M2.1, and M2.5 within 3.5 months, with progress on SWE-Bench Verified outpacing peers like Claude and GPT families.
- **MiniMax Agent Deployment:** M2.5 powers the MiniMax Agent platform, where users have built over 10,000 'Experts' for reusable workflows, and 30% of MiniMax's internal tasks are autonomously completed by M2.5.
- **Open Weights and Deployment:** Model weights are on Hugging Face under a Modified-MIT license, with deployment guides for SGLang, vLLM, and Transformers.