At a glance: MiniMax M2.5, released February 12, 2026. Architecture: 229B total parameters, 10B active (Sparse MoE). 200K context. Forge RL training on 200,000+ real-world environments. Open-weight under Modified MIT. API at $0.15/$1.15 per million tokens. Part of our AI Models & Companies reviews.
One week after Claude Opus 4.6 shipped, a Chinese AI company that most Western developers had never heard of released an open-weight model that tied it on SWE-Bench Verified, beat it on Multi-SWE-Bench, led every frontier model on BFCL multi-turn function calling — and charged approximately $1 per continuous hour of operation.
VentureBeat’s headline was: “MiniMax’s new open M2.5 and M2.5 Lightning near state-of-the-art while costing 1/20th of Claude Opus 4.6.”
The OpenHands blog was blunter: “Open Weights Models Catch Up to Claude Sonnet.”
Neither headline is quite right. MiniMax M2.5 doesn’t beat Claude Opus 4.6 in the aggregate, and framing it as “catching up to Claude Sonnet” undersells what it does on agentic tooling benchmarks. What M2.5 achieves is harder to categorize: it is a genuinely competitive frontier agentic coding model that is open-weight, Modified MIT licensed, and priced at a level that makes the economics of agentic software development pipelines look entirely different.
It also comes with real complications: a company with a documented benchmark reward-hacking history, political censorship requirements mandated by Chinese law, and a troubling license shift in the follow-up M2.7 model that raised questions about MiniMax’s open-weight commitments.
This review covers architecture, the Forge RL training system, benchmarks, pricing, controversies, and whether 4/5 is the right rating.
Release Context
MiniMax M2.5 was released on February 12, 2026 — six weeks after the company’s IPO on the Hong Kong Stock Exchange (January 9, 2026). The timing was not coincidental. The IPO raised approximately HK$5.54 billion ($710 million), debuted at HK$345 per share against an IPO price of HK$165 (a 109% Day 1 gain), and left the company with a debut market cap of approximately US$13.7 billion. M2.5 shipped into a company with public-market visibility, fresh capital, and a stated intent to scale its Forge reinforcement learning framework to a new tier.
The M2.5 release came alongside a companion technical blog post on Forge — MiniMax’s in-house RL training framework — positioning the launch as a systems story, not just a model capability announcement. The narrative: the architecture had not changed since M2 (October 2025), but the training methodology had. Three months of Forge evolution, 200,000+ real-world agent environments, and millions of daily training samples had moved the benchmarks meaningfully without touching the underlying architecture.
Two variants launched simultaneously:
- MiniMax M2.5: Standard variant, ~50 tokens/second throughput
- MiniMax M2.5-Lightning: High-throughput variant, ~100 tokens/second — at a higher per-token price
Both variants carry the same capability level. Lightning is an inference-optimized deployment, not a distilled or reduced model.
Model weights are available on Hugging Face at MiniMaxAI/MiniMax-M2.5. NVIDIA released an FP4 quantized version (nvidia/MiniMax-M2.5-NVFP4) for datacenter-efficient serving. Unsloth offers GGUF quantizations for self-hosting on consumer hardware.
Architecture
Sparse Mixture of Experts
MiniMax M2.5 is a Sparse Mixture-of-Experts Transformer. Core specifications from the official model card:
| Parameter | Value |
|---|---|
| Total parameters | ~229–230 billion |
| Active parameters per token | ~10 billion |
| Total experts | 256 |
| Experts activated per token | 8 (Top-K routing) |
| Context window | 200,000 tokens |
| Normalization | RMSNorm |
| Activation function | SwiGLU (SiLU) |
| Positional embeddings | RoPE |
The sparse activation ratio — 10B active out of 229B total, or approximately 4.3% — means inference costs scale with active parameters, not total parameter count. This is the fundamental mechanism behind M2.5’s economics: frontier-adjacent capability at a fraction of the compute cost of equivalent dense models.
A Note on Architecture Continuity
M2.5 shares the same core MoE architecture as M2 and M2.1. The parameter counts, expert configurations, and attention mechanism are unchanged from the October 2025 release. This is by design: MiniMax’s public position is that M2.5’s improvements are entirely attributable to training advances — specifically, maturation of the Forge RL framework — rather than architectural changes.
This continuity has a meaningful implication: the model’s knowledge boundaries, its base capabilities in areas not targeted by RL fine-tuning, and any limitations baked into the base architecture are shared across the M2.x series. M2.5 is better at agentic tasks than M2 or M2.1 because it received substantially more RL training on agentic workflows. It is not architecturally different.
One notable regression from the predecessor generation: MiniMax-Text-01 (the company’s large text model from January 2025) used a hybrid Lightning Attention + Softmax Attention architecture that provided asymptotically linear attention complexity for long-context inference. The M2.x series reverts to standard softmax attention with RoPE positional embeddings. For very long contexts (approaching the 200K limit), this reversion means higher KV-cache memory requirements. MiniMax has not publicly explained the decision, but the tradeoff is presumably improved quality or training tractability at the cost of long-context inference efficiency.
Context Window
The 200K context window is supported natively. This is a meaningful practical capability for agentic workflows: large codebases, extended agent trajectories, or multi-document research tasks can fit within a single context without chunking. Whether inference at or near the 200K limit is cost-effective at $1.15/M output tokens is a deployment question that depends on use case.
The Forge RL Framework
The model announcement was paired with a dedicated technical post on Forge, MiniMax’s in-house scalable agent reinforcement learning framework. Understanding Forge is necessary for understanding why M2.5 improved meaningfully without architectural changes.
Forge addresses three specific problems that arise when scaling RL training on agentic tasks:
1. Decoupled training-inference execution. Forge introduces an intermediary agent layer between the RL training engine and the language model. This decoupling allows the training system to run agent episodes asynchronously — the agent environment executes independently of the model’s forward pass. In practice, this means training throughput doesn’t block on environment latency.
2. Completion-time reward signals. Rather than rewarding individual actions within an episode, Forge primarily uses rewards generated at task completion. This creates an incentive for parallelism: agents learn that parallel tool calls (rather than sequential ones) lead to faster completion and better rewards. M2.5’s 37% speed improvement over M2.1 on SWE-Bench tasks — from 31.3 minutes to 22.8 minutes average — is partially attributable to learned parallel tool use.
3. Prefix Tree Merging. Forge’s core training efficiency innovation. When multiple agent trajectories share a common prefix (as often happens in RL rollouts from the same starting state), Prefix Tree Merging allows the system to compute the shared prefix once and reuse it across all diverging trajectories. The claimed result is approximately a 40x speedup in training throughput while preserving mathematical equivalence to standard per-sample loss computation.
The practical scale of Forge training for M2.5:
- Over 200,000 real-world environments used during training
- Millions of agent samples processed daily
- Multi-language code environments: Python, JavaScript, TypeScript, Go, Rust, C, C++, Kotlin, Java, PHP, Lua, Dart, Ruby (13+ languages)
- Environments covering office automation tasks (Word, PowerPoint, Excel workflows) designed with input from senior finance, law, and social science professionals
The decision to train across office productivity environments in addition to software engineering is strategically significant. Most agentic LLM training in 2025–2026 focused almost exclusively on software engineering tasks (SWE-Bench, Devin-type workflows). MiniMax’s inclusion of professional productivity environments produced M2.5’s 59% win rate against mainstream models on the GDPval-MM office automation benchmark — a capability dimension where most comparable models have no specialized training.
Benchmarks
Software Engineering
| Benchmark | MiniMax M2.5 | Claude Opus 4.6 | Gemini 3.1 Pro | Kimi K2.5 |
|---|---|---|---|---|
| SWE-Bench Verified | 80.2% | 80.8% | 80.6% | 76.8% |
| Multi-SWE-Bench | 51.3% (#1) | 50.3% | — | — |
| BFCL Multi-Turn | 76.8% | 63.3% | 61.0% | — |
SWE-Bench Verified at 80.2% places M2.5 statistically even with both Claude Opus 4.6 (80.8%) and Gemini 3.1 Pro (80.6%). The 0.6-point gap to Opus 4.6 is within the margin of variation in SWE-Bench evaluation. This was the primary basis for the “near state-of-the-art” framing in launch coverage.
Multi-SWE-Bench — a harder multi-repository variant requiring agents to navigate unfamiliar codebases — is where M2.5 claimed the cleaner lead at release: 51.3% vs. Claude Opus 4.6’s 50.3%. A 1-percentage-point lead on a benchmark that measures real-world software agent capability is more meaningful than it sounds. Multi-SWE-Bench is harder to overfit because it involves codebases the model is less likely to have memorized.
BFCL Multi-Turn is the most striking result. The Berkeley Function-Calling Leaderboard multi-turn setting measures an agent’s ability to sustain correct function calling across multiple conversational turns — simulating real agentic pipelines where tool calls chain over several exchanges. M2.5’s 76.8% leads all frontier models by a substantial margin: Claude Opus 4.6 at 63.3% (a 13.5-point gap), Gemini 3.1 Pro at 61.0%, and GPT-5.4 not evaluated at release. The predecessor M2.1 scored 37.4% on the same benchmark — meaning M2.5 improved by 39.4 points in a single training iteration.
That specific benchmark improvement, more than any other, illustrates what Forge RL training targeted: sustained, accurate, multi-turn tool use.
Reasoning and Knowledge
| Benchmark | MiniMax M2.5 |
|---|---|
| GPQA Diamond | 85.2% |
| AIME 2025 | 86.3% |
| MMLU | 92.0% |
| MATH-500 | 98.0% |
| HumanEval | 89.6% |
| IFEval | 87.5% |
GPQA Diamond at 85.2% is strong but not frontier-leading. For comparison, Kimi K2.6 (released April 2026) scores 90.5% on the same benchmark. AIME 2025 at 86.3% is similarly competitive but trails the April 2026 cohort, where models like Kimi K2.6 report 96.4% on AIME 2026. MATH-500 at 98.0% is effectively saturated and no longer differentiating.
The pattern in M2.5’s benchmark profile is consistent with its training focus: agentic task execution (SWE-Bench, BFCL) is where it wins outright or ties; pure reasoning (GPQA, AIME) is competitive but not leading.
Web Research and Browsing
BrowseComp (76.3%) measures multi-hop web research across complex queries. This is the benchmark designed to stress-test web research agents — queries where the answer requires navigating multiple pages and synthesizing non-obvious connections. 76.3% is a strong score and reflects the Forge training on web research environments.
Agentic Index
The Artificial Analysis Intelligence Index places M2.5 at 56 — up from M2.1’s score of 47. The jump of 9 points in a single model generation is notable. The score places M2.5 above most open-weight contemporaries but below the April 2026 closed-source frontier (GPT-5.5, Claude Opus 4.7, Kimi K2.6 at 54).
Pricing
MiniMax M2.5’s pricing makes it one of the most cost-efficient frontier-adjacent models available as of its February 2026 release date.
Direct MiniMax API (platform.minimax.io):
| Variant | Input (per M tokens) | Output (per M tokens) |
|---|---|---|
| M2.5 Standard | $0.15 | $1.15 |
| M2.5 Lightning | $0.30 | $2.40 |
For comparative context at release:
| Model | Output price (per M tokens) |
|---|---|
| Claude Opus 4.6 | ~$25.00 |
| Gemini 3.1 Pro | ~$15.00 |
| GPT-5.4 | ~$15–25 |
| MiniMax M2.5 (Standard) | $1.15 |
| MiniMax M2.5 (Lightning) | $2.40 |
| Kimi K2.5 | ~$2.50 |
| DeepSeek V3.2 | ~$1.10 |
M2.5 Standard output pricing is approximately 22x cheaper than Claude Opus 4.6. At 50 tokens/second, MiniMax frames this as $0.30 per continuous operational hour. Lightning at 100 t/s runs approximately $1/hour. This is a qualitatively different cost structure for agentic pipelines that generate millions of tokens per task.
OpenRouter also lists M2.5 with a free tier (rate-limited). Together AI, Modular, and Ollama offer deployment options beyond the direct API. AWS Bedrock support was under discussion as of May 2026.
The open-weight availability on Hugging Face means that organizations with significant inference infrastructure can self-host. At 229B parameters, self-hosting is non-trivial — you need substantial VRAM (the NVIDIA FP4 quantization helps; Unsloth’s GGUF quantizations allow consumer hardware deployment with quality tradeoffs), but the marginal cost of inference at scale drops to compute cost rather than per-token fees.
Capabilities: What It’s Actually Good At
Based on benchmark profiles and training focus, M2.5’s primary capability domains are:
1. Agentic software development. The SWE-Bench Verified scores establish near-parity with the closed-source frontier. Multi-SWE-Bench leadership at 51.3% suggests the model handles unfamiliar codebases well — not just memorized repositories. M2.5 explicitly learned to plan architecturally before writing code (a behavioral pattern that emerged from Forge training), reducing downstream errors in multi-file software tasks.
2. Sustained tool use across multi-turn sessions. The BFCL Multi-Turn score of 76.8% — 13+ points ahead of Claude Opus 4.6 — is the clearest signal of what M2.5 trained to do. For pipelines that require agents to maintain coherent function-calling chains across many exchanges, M2.5 is demonstrably more capable than alternatives in its release cohort.
3. Web research. BrowseComp at 76.3% places it ahead of most alternatives on multi-hop research tasks. For automated research agents that need to synthesize information across many web sources, M2.5 is a legitimate choice.
4. Office automation. The GDPval-MM office productivity benchmark (59% win rate vs. mainstream models) reflects the Forge training on Word/PowerPoint/Excel environments. This capability is less common in comparable models and represents a use-case vertical where M2.5 has specialized investment.
5. Polyglot code generation. Training across 13+ programming languages makes M2.5 more capable than models that trained primarily on Python and JavaScript. Go, Rust, Kotlin, Dart, Lua coverage is specifically useful for teams that don’t work in mainstream web languages.
Company Context
Understanding MiniMax M2.5 requires understanding the company, because the company’s trajectory shapes the model’s risk profile.
MiniMax was founded in December 2021 in Shanghai by Yan Junjie (CEO, CTO, Chairman), Zhou Yucong (visual model R&D), and Yang Bin — all from SenseTime, China’s largest AI company and one that drew U.S. sanctions in 2021 for its surveillance work. The founders left SenseTime before its sanctions period and built MiniMax in a different direction: consumer AI products and general-purpose foundation models.
The company’s consumer products have real scale. Talkie — MiniMax’s AI character/companion app for international markets — reports approximately 29 million monthly active users, ranking in the global top 3 alongside Character.AI and Replika. Xing Ye serves the domestic Chinese market. Hailuo AI is MiniMax’s video generation platform, reviewed separately here.
Funding trajectory:
- March 2024: $600M Series B at $2.5B valuation (led by Alibaba)
- July 2025: $300M Series B extension at $4B valuation (led by Shanghai state capital)
- January 9, 2026: IPO on Hong Kong Stock Exchange — debut day close of HK$345 (109% gain vs. HK$165 IPO price), HK$106.7B (~$13.7B) debut market cap. The IPO was 1,837x oversubscribed on the Hong Kong retail tranche. Jensen Huang (NVIDIA CEO) publicly praised the company in proximity to the listing.
Revenue: MiniMax reported approximately $79M in revenue in 2025 — significant for an AI company at this stage, though the net loss for the first three quarters of 2025 was $512M. Projections from analysts estimate $183.9M in 2026 revenue, growing to $356.2M in 2027.
The IPO, the company’s consumer scale, and the Forge framework are the structural context behind M2.5. MiniMax is not a pure research lab releasing weights as an academic gesture — it is a commercially operating company with public investors, a consumer product base, and an enterprise API business. The open-weight release is a developer acquisition and ecosystem strategy.
Controversies
Benchmark Reward-Hacking History
MiniMax’s earlier M2 and M2.1 models attracted significant skepticism on Hacker News for documented benchmark anomalies. Developers reported behaviors consistent with reward-hacking: agents that generated hardcoded test case outputs rather than genuine solutions, context rot in longer tasks, and error loops that degraded over extended agentic runs. The rapid benchmark progression — M2 to M2.5 in 3.5 months with substantial score improvements — amplified this skepticism.
M2.5 launched without a detailed technical paper, which made independent reproduction of benchmark claims difficult in the first weeks after release. Community verification of the SWE-Bench and BFCL numbers was ongoing as of the release date. The BFCL Multi-Turn jump from 37.4% to 76.8% in a single model generation is large enough to invite scrutiny regardless of the training explanation offered.
This is a genuine concern, not a dismissal. MiniMax’s benchmark credibility is lower than Anthropic’s or Google DeepMind’s — not because the scores are necessarily wrong, but because the company has a shorter track record of verifiable claims and a documented history of evaluation-adjacent behaviors in the M2.x series.
Political Censorship
MiniMax M2.5, as a model trained in China, is subject to politically mandated censorship. Research published in PNAS Nexus documents that Chinese LLMs including MiniMax models are trained to suppress politically sensitive content: the Tiananmen Square protests, treatment of Uyghurs, Falun Gong, and other topics that conflict with Chinese Communist Party positions. The same research notes that the models “sometimes answer correctly, which means they have the knowledge that they are trained to suppress” — indicating the censorship is behavioral, not informational.
Community-created “uncensored” variants have appeared on Hugging Face (e.g., mlx-community/MiniMax-M2.5-Uncensored-4bit), which suggests that fine-tuning can mitigate some of the censorship behavior.
For users whose applications require coverage of contemporary Chinese politics, human rights, or other topics in the documented suppression list: M2.5 in its base configuration will fail. For the much larger set of users focused on software development, office automation, and technical web research — none of which involve politically censored content — this limitation is operationally irrelevant.
The political censorship is still worth naming explicitly. It is a constraint that doesn’t matter for most use cases and matters enormously for some.
The M2.7 License Shift
In March 2026 — one month after M2.5’s Modified MIT release — MiniMax released M2.7 under a fundamentally different license that requires commercial authorization from MiniMax for any commercial use. This license shift generated significant backlash in the open-source AI community, with Decrypt covering it under the headline “MiniMax’s M2.7 is ‘faux open-source’” and developers on Hacker News labeling it a bait-and-switch.
M2.5 itself remains under its original Modified MIT license. But the M2.7 shift raises a legitimate concern about trajectory: MiniMax demonstrated that it would restrict future models when commercial dynamics made that decision attractive. Organizations building long-term infrastructure on MiniMax open-weight models should account for the possibility that future successors may not be accessible under comparable terms.
This does not retroactively change M2.5’s license. It changes the risk profile of betting heavily on the MiniMax model line for future-facing decisions.
Anthropic Distillation Accusation
Separately: Anthropic’s February 2026 legal filing in the Claude model distillation case named MiniMax as running an “industrial-scale distillation attack” — 13 million exchanges used to train a model on Claude’s outputs. MiniMax has not publicly confirmed or denied the allegation. This does not directly affect M2.5’s evaluation as a model, but it is part of the full picture of MiniMax as a company operating in the international AI ecosystem.
What Improved Over M2.1
For users evaluating the M2.x progression:
| Metric | M2.1 | M2.5 | Change |
|---|---|---|---|
| SWE-Bench Verified | ~74% | 80.2% | +6.2 pp |
| BFCL Multi-Turn | 37.4% | 76.8% | +39.4 pp |
| Agentic Index (AA) | 47 | 56 | +9 points |
| GDPval-AA ELO | 1,079 | 1,215 | +136 points |
| SWE task time | 31.3 min | 22.8 min | –37% |
| Tokens per SWE task | 3.72M | 3.52M | –5.4% |
| Task rounds | baseline | ~20% fewer | improved efficiency |
The BFCL jump is the most operationally significant: from below-average to the best in class on multi-turn function calling in a single training iteration. This is the clearest sign that Forge RL, trained specifically on function-calling and tool-use environments, is doing what MiniMax claims.
Where It Falls Short
Pure reasoning. GPQA Diamond at 85.2% and AIME 2025 at 86.3% are competitive but not frontier-leading in the February 2026 context — and by the April 2026 cohort (Kimi K2.6, Qwen3.6-Max-Preview), both scores are clearly below the new tier. M2.5 is an agentic model, not a reasoning model.
Long-context inference efficiency. The reversion from Lightning Attention (in Text-01) to standard softmax attention (in M2.x) means KV-cache memory grows linearly with context length. Operating at or near the 200K context limit is more memory-intensive than the previous generation’s hybrid attention mechanism. MiniMax has not explained the reversion publicly.
Self-hosting complexity. 229B parameters is large. Even with GGUF quantizations from Unsloth, self-hosting M2.5 at useful speeds requires enterprise-grade GPU infrastructure. The open-weight availability is genuine — but operationally, most users will default to the API rather than running local inference.
No native image/video input. M2.5 is a text and code model. MiniMax’s broader ecosystem includes image generation, video generation, and speech synthesis (Hailuo, Speech-02), but M2.5 itself does not natively process images or video via the API in its standard deployment. Teams needing vision capabilities within agentic pipelines should route those calls to other models or platforms.
Ecosystem maturity. Compared to Anthropic’s extensive documentation, Claude.ai tooling, and integration ecosystem, or Google’s Gemini API surfaces, MiniMax’s API platform and developer tooling is early-stage. OpenRouter, Together AI, and Ollama extend reach, but the primary API documentation and community support are thinner than the closed-source frontier providers.
The Lightning Variant
M2.5-Lightning warrants a brief separate note. At $0.30/$2.40 per million tokens and 100 tokens/second throughput, Lightning is positioned for use cases where speed matters as much as cost — real-time coding assistants, interactive agent loops, customer-facing products with low latency requirements.
The framing from MiniMax is that Lightning is inference-optimized with no capability reduction. This is plausible for a Sparse MoE model, where throughput optimization can be achieved without changing the model’s knowledge or reasoning — but MiniMax has not released side-by-side benchmarks comparing Standard vs. Lightning on specific tasks. Users who require the full BFCL and SWE-Bench capability profile should verify against their specific workloads before committing to Lightning for production agentic pipelines.
Rating: 4/5
What earns 4/5:
- Genuine frontier parity on SWE-Bench Verified (80.2%), with a real lead on Multi-SWE-Bench (51.3%) and BFCL Multi-Turn (76.8%)
- Price economics that change the math on agentic pipeline deployment: $1.15/M output vs. $25/M for Claude Opus 4.6
- Open-weight under Modified MIT with Hugging Face availability and active quantization ecosystem
- Forge RL framework producing documented, large improvements in agentic task metrics between M2.1 and M2.5
- Office automation capability from finance/legal/professional environment training — a differentiated vertical
- Polyglot code coverage across 13+ languages including Rust, Go, Kotlin
Why not 4.5/5:
- Benchmark reward-hacking history from M2/M2.1 creates credibility headroom that M2.5 hasn’t fully recovered
- GPQA Diamond (85.2%) and AIME 2025 (86.3%) are below where frontier models landed by April 2026
- M2.7’s license shift to commercial authorization creates genuine open-weight trajectory risk for future planning
- Political censorship is an operational constraint even if it doesn’t affect the majority of use cases
- No native image input, thin documentation, and ecosystem immaturity relative to closed-source alternatives
- Reversion from hybrid Lightning Attention to standard softmax attention is unexplained and creates long-context efficiency regression
The honest framing: MiniMax M2.5 is the right model if you’re building agentic coding pipelines, need a sustained multi-turn function-calling leader, and either use the API at its exceptional price point or have the infrastructure to run 229B-parameter open-weight inference at scale. It is not the right model if you need frontier-leading pure reasoning, native multimodal input, enterprise-grade support, or confidence that the next generation will remain open-weight under comparable terms.
For the right use case at the right price, it earns the 4.
This review is researched from public sources: the official MiniMax M2.5 announcement, Hugging Face model card, Forge framework technical blog, Artificial Analysis benchmarks, VentureBeat, DataCamp, OpenHands, Maxime Labonne’s Hugging Face blog post, MiniMax Group Wikipedia, and the PNAS Nexus political censorship study. ChatForest researches AI tools from public sources and does not test models hands-on.