Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchMiniMax released M2.5 — a 230 billion parameter mixture-of-experts model with 80.2% on SWE-Bench Verified, matching Claude Opus 4.6 speed at $0.29/M input tokens, trained across 200,000 real-world coding environments.
AI, startups & growth insights. No spam.
TL;DR: MiniMax released M2.5 on February 12, 2026 — a 230 billion parameter Mixture-of-Experts model that scores 80.2% on SWE-Bench Verified, placing it directly alongside Claude Opus 4.6 in coding performance. It runs at $0.29 per million input tokens, roughly 75% cheaper than Claude Opus 4.6's pricing, and completes coding tasks 37% faster than its predecessor M2.1. The model is open-weight, available on Hugging Face, and now hosted via Together AI and NVIDIA NIM for API access.
MiniMax launched M2.5 on February 12, 2026, initially with limited public coverage. By early March, the model began trending widely across developer communities as benchmark analyses from Artificial Analysis and others confirmed its competitive position against the top US frontier models.
The headline architecture numbers: 230 billion total parameters with 10 billion active per forward pass. That ratio — 10B active from a 230B pool — is the defining characteristic of Mixture-of-Experts design, and it is why M2.5 can match much larger dense models on compute-intensive benchmarks while remaining cost-competitive for API deployment.
Context window: 196,600 tokens. That covers approximately 150,000 words of text, enough to load a substantial codebase and its full test suite in a single context. For software engineering tasks in particular, this matters because the model can hold all relevant files, documentation, and error traces simultaneously without truncation.
The benchmark headline: 80.2% on SWE-Bench Verified. SWE-Bench Verified is the primary standard for evaluating AI coding assistants on real GitHub issues, and 80.2% puts M2.5 in the same tier as Claude Opus 4 and Sonnet 4, which Anthropic positioned as frontier-grade coding models when they launched with their own SWE-Bench records. M2.5 also posts 51.3% on Multi-SWE-Bench, a multi-repository variant of the benchmark that tests cross-codebase reasoning, and 76.3% on BrowseComp, a web browsing and comprehension evaluation.
Speed: MiniMax reports that M2.5 completes SWE-Bench tasks 37% faster than M2.1. Independently, Artificial Analysis found M2.5's response latency comparable to Claude Opus 4.6 in real-world API tests, which is notable given M2.5 costs substantially less per token.
The Mixture-of-Experts (MoE) pattern solves a fundamental problem in large language model scaling: as you add parameters to improve capability, inference cost scales proportionally — unless you route each input to only a subset of the network.
M2.5 uses 230 billion parameters distributed across a large number of expert sub-networks. For any given input, a learned routing mechanism selects which experts to activate. The result is that each forward pass uses roughly 10 billion parameters of compute, even though the total model has 230 billion available. The model effectively becomes more capable as the expert count grows, but the per-token inference cost tracks with the active parameter count, not the total.
This is not unique to MiniMax. Mixtral 8x7B popularized the approach in the open-weight space. Google's Gemini 1.5 and Gemini 3.1 Pro use MoE internally. What distinguishes M2.5 is the combination of the 23:1 total-to-active ratio with the specific training corpus designed around software engineering tasks.
The practical implication for API pricing is direct. Running a dense 70B model requires 70B parameter operations per token. Running M2.5 requires approximately 10B, even though the model has access to 230B parameter worth of learned specializations. Infrastructure providers can pass that efficiency to customers through lower token prices, which is reflected in M2.5's $0.29 per million input tokens.
One trade-off MoE architectures accept: memory footprint. Hosting the full 230B parameter model requires significant GPU memory even when only 10B is active per pass. This is why smaller organizations may find M2.5 impractical to self-host, making Together AI and NVIDIA NIM's hosted API endpoints particularly relevant.
SWE-Bench Verified is the test that matters most for evaluating M2.5's positioning. The benchmark presents models with real GitHub issues from popular open-source repositories — the task is to write code that, when applied as a patch, causes previously failing tests to pass. "Verified" refers to a curated subset where human annotators have confirmed each issue is valid and unambiguous. Scoring 80.2% means M2.5 correctly resolves four out of five such issues.
For context, when Claude Opus 4 and Sonnet 4 launched, Anthropic highlighted SWE-Bench performance as a key differentiator. M2.5 achieving 80.2% without additional scaffolding — that is, running the model directly without agentic wrappers — places it in competitive range with those models.
Multi-SWE-Bench (51.3%) extends the single-repository format to multi-repository scenarios: the model must understand how changes in one codebase affect dependent codebases. This tests whether the model can reason about dependency graphs and API contracts across projects, not just make local fixes. The benchmark is significantly harder, and 51.3% represents a strong result among models that have published scores on it.
BrowseComp (76.3%) evaluates a different capability: the model's ability to browse web content, extract information, and answer compound research questions accurately. This is relevant for agentic use cases where the model must retrieve current information, navigate multi-step research tasks, and synthesize findings. OpenAI's deep research products have driven interest in this benchmark category, and 76.3% positions M2.5 as competitive for research-augmented workflows.
One benchmark MiniMax does not prominently highlight: pure reasoning evaluations like MATH or competition-grade problem sets. The model's design emphasis on software engineering and coding tasks is intentional, and enterprises evaluating M2.5 for mathematical or scientific reasoning should conduct independent assessment rather than assuming benchmark parity with models optimized for those domains.
The most architecturally distinctive aspect of M2.5 is not its parameter count or its MoE design — it is the pre-planning approach MiniMax calls "active project decomposition," for which they have filed a patent.
The problem it addresses is well-known to anyone who has observed frontier AI models attempt large coding tasks: models trained to predict the next token naturally want to begin writing code immediately. They generate a plausible-looking first function, then a second, then realize mid-way that the approach is incompatible with what the problem requires. The result is code that looks coherent at the local level but fails at the architectural level.
Active project decomposition inverts this sequence. Before M2.5 generates a single line of implementation code, it produces a structured project plan: component decomposition, interface definitions, dependency ordering, and a mapping of which parts of the codebase each sub-task will affect. This plan is then used to guide the implementation phase, much the way an experienced software architect produces a design document before handing off to developers.
This is distinct from chain-of-thought reasoning in an important way. Chain-of-thought prompting asks the model to show its reasoning step-by-step, which improves accuracy on reasoning tasks by making intermediate steps explicit. Active project decomposition is a structural approach to task sequencing: the model is not just thinking aloud about a single problem — it is decomposing a project into a dependency-ordered work breakdown structure before any code generation begins.
The practical result shows up in the benchmark numbers. SWE-Bench Verified issues often require touching multiple files and understanding how components interact. Models that start writing code immediately tend to make locally sensible but globally inconsistent changes. M2.5's planning pass reduces this class of error.
The extended thinking capability — chain-of-thought reasoning for complex problems — layers on top of the planning approach for tasks where step-by-step reasoning is additionally useful. For multi-step debugging or algorithmic problems, extended thinking and project decomposition work in combination.
The pricing gap between M2.5 and US frontier models is the clearest reason for enterprise interest. Here is the direct comparison across models competitive on SWE-Bench:
| Model | SWE-Bench Verified | Input price ($/M tokens) | Output price ($/M tokens) |
|---|---|---|---|
| MiniMax M2.5 | 80.2% | $0.29 | $1.20 |
| Claude Opus 4.6 | ~80%+ | $1.20 | $6.00 |
| GPT-5.4 | competitive | $1.50+ | $6.00+ |
| Gemini 3.1 Pro | competitive | $1.25 | $5.00 |
On input tokens, M2.5 is approximately 75% cheaper than Claude Opus 4.6 and GPT-5.4. On output tokens, the gap narrows but remains substantial: M2.5 at $1.20/M output versus Claude Opus 4.6 at $6.00/M is an 80% reduction.
For workloads dominated by coding tasks — automated PR review, code generation pipelines, large-scale refactoring — this difference compounds quickly. An organization running 10 billion input tokens per month through Claude Opus 4.6 pays approximately $12,000 in input token costs alone. The same workload through M2.5 costs $2,900. At scale, the savings fund substantial additional AI usage or simply reduce operating costs.
The important caveats for this comparison: pricing can change, and total cost of ownership includes factors beyond token price — integration complexity, latency characteristics, failure modes, and how well the model handles edge cases in your specific domain. The cost arbitrage is real, but the decision to switch requires validating performance on your actual workload, not just benchmark equivalence.
MiniMax's technical documentation states that M2.5 was trained across more than 200,000 real-world coding environments. This figure refers to distinct development environments — project structures, build systems, test suites, dependency configurations — rather than individual code samples.
The distinction matters because training on diverse environments teaches the model to navigate the messiness of real codebases: inconsistent conventions, missing documentation, legacy code patterns, and environment-specific tooling. A model trained only on curated code samples from GitHub may produce clean code in isolation but fail when asked to work within an existing codebase that uses unconventional patterns.
The 13 supported programming languages reflect the breadth of the training corpus: Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby. This covers the majority of enterprise application development surface area. Rust and Kotlin specifically suggest attention to systems programming and Android development respectively, which are areas where code generation quality varies significantly across models.
Python and JavaScript support is table stakes for any frontier coding model. The inclusion of Lua is notable — Lua is widely used in game development (Roblox, World of Warcraft addons) and embedded scripting (Redis, nginx), domains where AI coding assistance has historically been underserved by frontier models.
M2.5 is available through three primary access points as of March 2026.
Hugging Face (huggingface.co/MiniMaxAI/MiniMax-M2.5) hosts the model weights directly. This is the open-weight release, meaning organizations with sufficient infrastructure can download and self-host. The weights are available under a license that permits commercial use, though MiniMax's specific terms should be reviewed before enterprise deployment. Self-hosting the full 230B parameter model requires substantial GPU infrastructure — at minimum, multiple high-memory GPUs — which limits this path to organizations with existing large-scale inference infrastructure.
Together AI's model page provides hosted API access using OpenAI-compatible endpoints. For teams already building on the OpenAI API format, switching to M2.5 via Together AI requires minimal code changes — typically a base URL swap and model name update. Together AI's infrastructure handles the hosting complexity, and pricing follows M2.5's published rates.
NVIDIA NIM (build.nvidia.com/minimaxai/minimax-m2.5) offers another hosted API path, optimized for NVIDIA GPU infrastructure. NIM endpoints are designed for enterprise deployments that require SLA guarantees and integrate with NVIDIA's broader AI infrastructure stack. For organizations running on NVIDIA hardware or using other NVIDIA AI services, NIM provides a coherent integration path.
The combination of open weights and multiple hosted API options positions M2.5 differently from closed models like Claude Opus 4.6 and GPT-5.4. Enterprises that require data residency control, on-premises deployment, or the ability to fine-tune on proprietary data have a viable path that does not exist with closed frontier models.
M2.5 did not appear in isolation. March 2026 has produced a notable cluster of major model releases from Chinese AI labs, and understanding M2.5's positioning requires that context.
Tencent, Alibaba, Baidu, and ByteDance have each released or announced significant model updates in the same window. The simultaneous activity reflects several converging pressures: access to competitive GPU infrastructure through Huawei (a dynamic that DeepSeek V4's hardware decisions have accelerated), growing domestic demand for frontier AI capabilities, and intensifying competition for enterprise customers who are actively evaluating non-US alternatives.
MiniMax's financial position provides context for their R&D capacity. The company reported 159% year-over-year revenue growth in their 2025 financial results. That growth rate, applied to a company already operating at meaningful scale, implies substantial resources available for model development and infrastructure investment. Unlike some earlier Chinese AI labs that released impressive research models without commercial traction, MiniMax appears to be growing its revenue base fast enough to sustain continuous frontier model development.
The broader pattern is that Chinese AI labs are narrowing the gap with US frontier models faster than most Western observers expected, across both raw benchmark performance and practical deployment capabilities. M2.5's combination of competitive SWE-Bench scores, open weights, and aggressive pricing reflects a deliberate strategy: capture enterprise workloads where cost sensitivity is high and US export restrictions have created customer anxiety about supply chain dependability.
Honest assessment of M2.5 requires acknowledging where it falls short of the claims implied by its headline benchmark.
First, the model is explicitly optimized for software engineering tasks. Its benchmark portfolio — SWE-Bench, Multi-SWE-Bench, BrowseComp — reflects this focus. Organizations evaluating M2.5 for advanced mathematical reasoning, scientific literature synthesis, or long-document summarization should not assume benchmark parity with models optimized for those tasks. The 80.2% SWE-Bench score does not generalize automatically.
Second, context reliability at the 196K token limit deserves scrutiny. Many models report maximum context windows that perform well at medium lengths but degrade significantly as inputs approach the maximum. MiniMax has not published needle-in-a-haystack or long-context retrieval benchmark results comparable to what Anthropic and Google have released for their models. Enterprise deployments that depend on reliable retrieval from very long contexts should validate this empirically.
Third, the agentic use case creates additional unknowns. SWE-Bench measures single-turn or scaffolded coding performance. Real agentic software engineering — the model running in a loop, using tools, executing code, reading error output, and revising — introduces failure modes that benchmark scores don't capture directly. MiniMax has not published detailed results from long-horizon agentic evaluations comparable to what Anthropic has released for Claude's computer use capabilities.
Fourth, the multi-language support list includes languages where quality may vary substantially. Dart and Lua support are listed, but neither has been prominently benchmarked. Developers working in these languages should evaluate code quality independently rather than assuming the same performance as Python and TypeScript.
The decision framework for enterprises evaluating M2.5 comes down to three questions: What is the primary workload? How cost-sensitive is the deployment? And what are the data residency requirements?
M2.5 makes the strongest case when the workload is primarily software engineering: code generation, automated review, test writing, refactoring, bug fixing. The model's training focus, pre-planning architecture, and SWE-Bench performance all point to this domain. If your AI spend is dominated by coding tasks, the 75% input token cost reduction relative to Claude Opus 4.6 represents a significant operating expense reduction that warrants serious evaluation.
Cost sensitivity matters most at scale. For small volumes — a few million tokens per month — the absolute dollar difference between M2.5 and Claude Opus 4.6 is modest. For large-scale pipelines processing billions of tokens monthly, the gap is material. The inflection point depends on your specific usage, but for most enterprise coding automation use cases, it falls somewhere between 100M and 1B tokens per month.
Data residency and supply chain considerations are increasingly relevant. Some enterprises have concerns about routing sensitive code through US-based API providers, either for compliance reasons or due to concerns about training data usage. Others have the reverse concern about routing code through Chinese-headquartered providers. Both are legitimate considerations. The open-weight availability of M2.5 is relevant here: self-hosting on your own infrastructure eliminates the data transmission concern entirely, at the cost of infrastructure complexity.
Where Claude Opus 4.6 and GPT-5.4 maintain clearer advantages: multimodal tasks requiring vision, domains outside software engineering where they have stronger benchmark coverage, mature agentic frameworks with proven reliability in production, and enterprise support structures that some organizations require. Claude Opus 4 and Sonnet 4's hybrid thinking capabilities also give Anthropic's models an edge on tasks requiring deep multi-step reasoning that goes beyond code generation.
The practical recommendation: treat M2.5 as a viable primary model for coding-focused deployments where you can validate performance on your actual workload before committing. Run a parallel evaluation against your current model on a representative sample of real tasks — not just public benchmarks — and let the results guide the decision. The cost gap is large enough that a modest performance equivalence justifies switching for cost-sensitive workloads, but not large enough to accept meaningful capability degradation in tasks your deployment depends on.
MiniMax M2.5 is the clearest signal yet that competitive frontier AI coding capability is no longer exclusive to a small group of US labs. The combination of open weights, aggressive pricing, and benchmark performance that legitimately challenges Claude Opus 4.6 creates a genuine alternative for enterprises building on AI-powered software development workflows. Whether that alternative is right for a specific deployment depends on careful workload evaluation — but the evaluation is now worth doing.
Sources: MiniMax official announcement (minimax.io/news/minimax-m25), Hugging Face model page (huggingface.co/MiniMaxAI/MiniMax-M2.5), The Information reporting on MiniMax M2.5 launch, Artificial Analysis MiniMax-M2.5 intelligence and performance analysis, NVIDIA NIM model endpoint (build.nvidia.com/minimaxai/minimax-m2.5), Together AI model page.
Meta has quietly revealed its Llama 4 lineup: Mango (flagship multimodal) and Avocado (coding-specialized) models launching H1 2026, with open weights that could democratize AI access.
DeepSeek V4 launches with 1 trillion total parameters, native multimodal capabilities, a 1M token context window, and pricing 50x cheaper than GPT-5.2 — all open-source and optimized for Huawei chips.
MiniMax posts $79M revenue with 159% year-over-year growth after its $614M Hong Kong IPO, with over 70% of revenue coming from outside China.