MiniMax M1 and M2.5: The Developer Guide to Open-Weight Models That Beat Claude on Benchmarks

Open-weight AI models have been chasing closed-model performance for years. MiniMax just caught up. Their M2.5 model — released under an MIT license with only 10 billion active parameters — scored 80.2% on SWE-Bench Verified, putting it within striking distance of Claude Opus 4.6 and ahead of GPT-5.3 Codex.

The cost? Roughly $0.25 per million input tokens. That is 20x cheaper than Claude Opus 4.6 and 5x cheaper than GPT-5.3. For developers running high-volume coding workloads, this changes the economics entirely.

This guide covers both MiniMax models (M1 and M2.5), how they compare to Claude, GPT, Gemini, and DeepSeek on real benchmarks, what the pricing actually means at scale, and a decision framework for choosing the right model for your workflow.

Who Is MiniMax?

MiniMax is a Shanghai-based AI company founded in 2021 by Yan Junjie, a former deputy director of Huawei's AI lab. The company raised $619 million in funding and IPO'd on the Hong Kong Stock Exchange in January 2026 at a $6.5 billion valuation. Investors include Alibaba, Tencent, and Goldman Sachs.

MiniMax is one of China's "AI Six Little Tigers" — a group of startups that emerged as serious competitors to both Chinese tech giants and Western AI labs. They operate Hailuo, a consumer AI platform with over 35 million monthly active users, and have been building foundation models since 2023.

For Western developers, the relevant context is that MiniMax is not a scrappy startup releasing a hobby project. They are a publicly traded company with billions in backing, and their models are trained on significant compute budgets. The open-weight releases (M1 and M2.5) represent a deliberate strategy to build ecosystem adoption outside China.

MiniMax M1: Architecture That Actually Matters

MiniMax M1 is a 456-billion parameter mixture-of-experts model with 45.9 billion active parameters per forward pass. The architecture combines standard transformer attention with a proprietary technique called Lightning Attention.

Lightning Attention is the interesting part. Standard transformer attention scales quadratically with sequence length — doubling your context window roughly quadruples the compute cost. Lightning Attention uses a hybrid approach: standard softmax attention for nearby tokens (where precision matters most) and a linear attention variant for distant tokens. MiniMax reports this reduces FLOPs to roughly 25% of what a model like DeepSeek R1 requires at 100K token generation length.

The practical result is a 1-million-token context window that actually works at reasonable speed. Many models advertise large context windows but degrade significantly in quality or speed as you approach the limit. MiniMax claims M1 maintains coherent recall across the full context.

M1 was also trained using CISPO (Constrained Iterative Self-Play Optimization), a reinforcement learning algorithm that constrains each training iteration to stay close to the previous policy. This is meant to avoid reward hacking and mode collapse — problems that plague standard RLHF at scale. MiniMax reports CISPO produces more stable behavior during extended reasoning chains.

MiniMax M2.5: The Open-Weight Model That Beats Claude on SWE-Bench

MiniMax M2.5 is the more practically relevant model for most developers. It is a 230-billion parameter MoE with only 10 billion active parameters — significantly smaller than M1 but achieving higher coding benchmark scores.

The headline number: 80.2% on SWE-Bench Verified. That puts M2.5 alongside Claude Opus 4.6 (~80%) and ahead of GPT-5.3 Codex (~75%). It is the first open-weight model to reach this tier.

M2.5 also scores 76.3% on BrowseComp (web browsing tasks) and 76.8% on BFCL (Berkeley Function Calling Leaderboard) — tool use. These are agent-relevant benchmarks that measure how well a model can interact with external tools and APIs, not just generate code in isolation.

The Lightning variant of M2.5 generates at roughly 100 tokens per second — fast enough for interactive coding sessions. The model is released under an MIT license, which means no restrictions on commercial use, modification, or redistribution.

Benchmark Comparison: M2.5 vs Claude vs GPT-5 vs Gemini

Six benchmarks that matter for real-world developer work. Note that not every model has published scores on every benchmark — we show only confirmed results.

Specs at a Glance

Pricing: 10-20x Cheaper Is Not a Typo

The pricing gap between MiniMax and closed models is the single most disruptive aspect of these releases.

To put this in concrete terms: a workload of 10 million input tokens and 2 million output tokens per day costs roughly $4.70/day with MiniMax M2.5 vs $100/day with Claude Opus 4.6. Over a month, that is $141 vs $3,000. Over a year, $51,000 vs $1,095,000.

This does not mean MiniMax is always the right choice. Cost per token is not cost per task. If Claude Opus 4.6 solves a complex architecture problem in one pass that MiniMax takes five attempts to get right, the effective cost flips. But for high-volume, well-defined coding tasks — test generation, code review, bug triage, documentation — the economics favor open-weight models heavily.

Self-hosting changes the math further. Running MiniMax M2.5 on your own GPUs eliminates per-token costs entirely. The tradeoff is upfront hardware investment (roughly $120K-$160K for 4x A100 80GB) and ongoing operational overhead. For teams processing millions of tokens per day, self-hosting pays for itself within months.

Context Windows and Long-Context Performance

Context window size determines how much code, documentation, or conversation history a model can process in a single call. Here is how the models compare.

MiniMax M1's 1-million-token context matches Gemini 2.5 Pro and dwarfs GPT-5.3's 400K. The Lightning Attention mechanism means M1 maintains reasonable speed at these lengths, whereas many models with large advertised context windows slow to a crawl in practice.

For most coding tasks, 128K-400K tokens is sufficient. The 1M window matters for specific use cases: analyzing entire monorepos, processing long legal or financial documents, or maintaining context across very long pair-programming sessions. If your workflow involves these scenarios, MiniMax M1 or Gemini 2.5 Pro are the current leaders.

Developer Experience: API, Self-Hosting, and Tooling

MiniMax provides an OpenAI-compatible API, which means most existing tooling works out of the box. You can use the standard OpenAI Python SDK by pointing the base URL to MiniMax's API endpoint.

For self-hosting, MiniMax M2.5 supports deployment via vLLM and SGLang — the two most popular open-source inference frameworks. The full model requires approximately 4x 80GB GPUs (A100 or H100). The Lightning variant with 10B active parameters can run on smaller hardware.

One practical consideration: ecosystem maturity. Claude and GPT have years of tooling, documentation, and community knowledge. MiniMax's ecosystem is younger. You will find fewer Stack Overflow answers, fewer tutorials, and less third-party tooling. For teams comfortable reading model cards and GitHub issues, this is a minor friction. For teams that rely on polished documentation and enterprise support, it matters more.

Integration with coding tools like Claude Code or Cursor is possible via the OpenAI-compatible API. Claude Code, for example, supports custom API endpoints — you can point it at a MiniMax-compatible server to use M2.5 as the backend model.

The Open-Weight Advantage (and Its Limits)

The MIT license on MiniMax M2.5 is genuinely permissive. Commercial use, modification, redistribution — all permitted without restriction. This is meaningful for several reasons:

• Data sovereignty: You can run the model entirely on your own infrastructure. No data leaves your network. This matters for regulated industries (healthcare, finance, government).
• Customization: You can fine-tune on your own codebase, proprietary documentation, or domain-specific data. Closed models offer limited fine-tuning options.
• Cost control: No per-token pricing. Once you have the hardware, marginal cost per query approaches zero.
• No vendor lock-in: You are not dependent on MiniMax's API availability, pricing changes, or terms of service updates.

The limits are equally real. Open-weight models lack the enterprise support infrastructure that Anthropic and OpenAI provide. There is no SLA, no dedicated account team, no compliance certifications (SOC 2, HIPAA BAA). If your organization requires these, closed models remain the safer choice for now.

Trust and provenance also matter. MiniMax is a Chinese company, and some organizations have policies restricting the use of models trained on undisclosed data by foreign entities. This is a business decision, not a technical one, but it is worth acknowledging.

Which Model Should You Use?

The answer depends on four factors: budget, privacy requirements, task complexity, and support needs. Here is a decision framework.

What This Means for the AI Market

MiniMax M2.5 reaching Claude-tier coding performance at 1/20th the price signals a structural shift. The gap between open-weight and closed models is shrinking faster than most predictions suggested. Six months ago, the best open models were 10-15% behind Claude on SWE-Bench. Today that gap is effectively zero.

This does not mean closed models are dead. Anthropic and OpenAI still lead on agentic capabilities, enterprise tooling, safety infrastructure, and support. For teams that need maximum capability with full support, Claude Opus 4.6 and GPT-5.3 remain the right choice. But for teams where cost or data sovereignty are primary constraints, the open-weight tier is now genuinely competitive.

The most practical takeaway: the era of single-model strategies is over. If you are looking for the best AI model for coding in 2026, the answer is not one model — it is a portfolio. The smartest development teams are routing different tasks to different models based on cost, capability, and privacy requirements. MiniMax M2.5 for high-volume coding. Claude Opus 4.6 for complex agentic workflows. GPT-5.3 for multimodal tasks. DeepSeek for general reasoning. As the best open-source (open-weight) models close the gap with closed alternatives, the economic case for a multi-model strategy only gets stronger.

Methodology and Sources

Benchmark data sourced from MiniMax (GitHub), Anthropic, OpenAI, SWE-Bench Verified leaderboard, Berkeley Function Calling Leaderboard (BFCL), and Artificial Analysis as of February 2026. Pricing reflects official API rates and third-party estimates where official pricing is not published. Where exact benchmark numbers are not publicly available, we note estimates based on independent testing and leaderboard positioning. The MiniMax M1 technical paper provides detailed architecture specifications for Lightning Attention. "Winner" designations reflect the current public data and may shift as more detailed benchmarks are published.