MiniMax (Text Models)

MiniMax is a Shanghai-based AI research company founded in early 2022, operating a lineup of large language models that compete directly with the most capable open-weight systems in the world. The lab's text model series, spanning MiniMax-Text-01, M1, M2, and the current M2.7 flagship, covers long-context reasoning, software engineering, and multi-step agentic task execution. This listing covers the text model API only, separate from MiniMax's Hailuo video platform and its audio and music generation products, which serve different use cases entirely.

The MiniMax text API is accessible at platform.minimax.io, via an Anthropic-compatible endpoint, or through OpenRouter. The lineup includes open-weight models under Apache 2.0 (M1, M2) and proprietary variants. Context windows range from 66K tokens on the M2-Her roleplay variant up to 1 million tokens on M1 and MiniMax-01, with a reported 4M-token extrapolation capability at inference. Pricing runs from $0.15 per million input tokens (M2.5) to $0.40 per million (M1), putting MiniMax among the least expensive frontier-class APIs for code and long-document work. Key use cases: agentic software pipelines, large-codebase analysis, multi-turn instruction following, and bulk content processing where cost per token matters.

MiniMax at a glance, May 2026

The current MiniMax text model lineup, from oldest to newest:

• MiniMax-Text-01 / MiniMax-01 (January 2025): 456 billion total parameters, 45.9 billion activated per token through a Mixture-of-Experts design. Introduced the "lightning attention" mechanism enabling 1M-token native context and 4M-token inference extrapolation. Benchmarks at launch matched GPT-4o and Claude 3.5 Sonnet while offering a context window 20-32x longer. Open-sourced on HuggingFace under Apache 2.0.

• MiniMax-M1 (June 16, 2025): 456B total parameters, hybrid-attention reasoning model, 1M context, up to 80K output tokens. Available as M1-40k and M1-80k variants. The full RL training run cost $534,700 across 512 H800 GPUs over three weeks, using MiniMax's proprietary CISPO reinforcement learning algorithm. SWE-bench scores: 55.6% (40k) and 56.0% (80k).

• MiniMax-M2 (October 23, 2025): 230B total parameters, 10B active per inference via MoE. Ranked first among all open-weight systems on the Artificial Analysis Intelligence Index at launch. LiveCodeBench ~83%, SWE-bench verified 69.4%. 197K context. Output speed 114.5 tokens/second.

• MiniMax-M2.1: Multilingual and multi-task coding focus. 1M context window. Targets multi-step programming tasks and enterprise language coverage.

• MiniMax-M2.5: SWE-bench 70.40%, AIME 88.75%. Lowest price in the lineup at $0.15/1M input tokens. Approximately 16x cheaper than Claude Opus 4.6 input rates for equivalent tasks.

• MiniMax-M2.7 (March 18, 2026): Current flagship. SWE-Pro 56.22%, VIBE-Pro 55.6%, Terminal Bench 2 57.0%. GDPval-AA ELO of 1495, highest among open-source models in that evaluation. 205K context.

• MiniMax-M2-Her: Specialized variant for roleplay and multi-character immersive interaction. 66K context.

Access options include direct HTTP, the OpenAI SDK, or MiniMax's Anthropic-compatible endpoint (https://api.minimax.io/anthropic), making it a drop-in for codebases already built on Anthropic tooling.

What MiniMax is actually good at

The M-series consistently leads or ties for first among open-weight models on software engineering benchmarks. At M1's June 2025 release, it led all open models on TAU-bench (agent tool use), outperforming Gemini 2.5 Pro and OpenAI o3 on long-context tasks. The M2 and M2.5 continued that trajectory, with M2.5 reaching 70.40% on SWE-bench verified, matching scores from models costing several times more per token.

The genuine differentiator is the combination of long context and low active parameter count. MiniMax M1 with its 1M-token window processes approximately 70% fewer compute resources than DeepSeek R1 does at 100K tokens, per MiniMax's own measurements. For teams doing large-codebase analysis, document review, or long-horizon agent loops, that math matters.

"Integrating the MiniMax M2 series into our platform has been a significant win for our users. We've found that M2.1 handles the nuances of complex, multi-step programming tasks with a level of consistency that is rare in this space." - developer platform team, quoted in MiniMax M2.1 launch coverage, January 2026

The open weights are a real advantage for teams that need to self-host or fine-tune. M1 and M2 are both Apache 2.0, allowing commercial use and modification. The MiniMax-01 paper (arXiv:2501.08313) provides full architectural details, making it one of the more transparent frontier-class model releases from any lab globally.

"We're excited for powerful open-source models like M2.1 that bring frontier performance, and in some cases exceed the frontier, for a wide variety of software development tasks. Developers deserve choice, and M2.1 provides that much-needed choice." - developer quoted in MiniMax M2.1 launch press, January 2026

Where MiniMax breaks: the failure modes users keep hitting

The rate limit system is the most frequently cited friction point. MiniMax uses fixed 5-hour windows rather than rolling rate limits. If you exhaust your allowance at 9:45 AM, you wait until 10:00 AM for the next fixed window to open. Developers accustomed to OpenAI or Anthropic's rolling windows find this disorienting, and the fixed windows interact poorly with bursty agent workloads. Users who created accounts after March 23, 2026 also face weekly quota caps that earlier users do not, creating unequal access tiers within the same product.

The models are verbose. Artificial Analysis measured M2 generating "70 million output tokens during evaluation, somewhat higher than average," which translates directly to elevated output costs in production. For summary or classification tasks with short expected outputs, the verbosity can meaningfully inflate bills.

Local inference is impractical for most teams. MiniMax-M1 at 456 billion parameters requires hardware most developers cannot access. Community discussions on HuggingFace from late 2025 raised uncertainty about llama.cpp compatibility. The M2 at 230B is more feasible but still out of reach without a cluster.

Content restrictions apply. Chinese regulatory requirements (a 2023 law prohibiting content "damaging to national unity and social harmony") mean the models filter political topics, particularly those touching on Chinese governance and territorial issues. For general coding and writing tasks this rarely surfaces, but it is a documented constraint that teams in news, policy research, or geopolitical analysis will encounter.

On pure mathematical reasoning, MiniMax M2 trails more focused reasoning models. AIME-25: MiniMax M2 at 78% versus Kimi K2 Thinking at 94-99%. If step-by-step math or formal proof generation is the primary use case, Kimi or Qwen3.5 is a better match.

MiniMax vs. Qwen vs. DeepSeek

All three are major Chinese open-weight labs releasing models in rapid succession through 2025-2026. The differences are architectural and operational, not just marketing.

MiniMax vs. Qwen (Alibaba): Both use Mixture-of-Experts with sparse activation. Qwen3.5 activates approximately 37 billion parameters per forward pass; MiniMax M2.7 activates 10 billion from 230 billion total, making per-token compute costs lower for MiniMax. Qwen3.5's context window reaches ~991K tokens, comparable to MiniMax M1's 1M. On competitive programming, Qwen3.5 leads (LiveCodeBench 85.33% vs. MiniMax M2's ~83%). On knowledge-intensive tasks, Qwen3.5 scores 87.37% on GPQA Diamond. Flagship input pricing is equivalent at $0.30/1M for both, though MiniMax M2.5 at $0.15 undercuts Qwen's older variants. The choice between them for most coding and analysis tasks is genuinely close; Qwen edges ahead on knowledge retrieval and competitive math, MiniMax edges ahead on context length and raw agentic tool use.

MiniMax vs. DeepSeek V3.2: The mechanical gap is throughput versus context depth. DeepSeek V3.2 generates approximately 230 tokens per second; MiniMax M2 generates 114.5. If speed is the primary constraint, DeepSeek is roughly twice as fast. When MiniMax M1 launched in June 2025, its 1M-token context was 8x larger than DeepSeek R1's 128K window, a concrete architectural gap. DeepSeek V3.2 has since extended context, but MiniMax M1 and MiniMax-01 remain the standard-setters for very long input handling. Rate limit design also differs: DeepSeek uses rolling windows, MiniMax uses fixed 5-hour windows. Teams with bursty usage patterns generally report DeepSeek's system is easier to work with.

For completeness, Kimi K2 from Moonshot AI is the third major competitor in this tier. Kimi uses 1 trillion total parameters (32B active), outperforms MiniMax M2 on BrowseComp (60.2% vs. 44.0%), and matches Claude Opus 4.6 on AIME math. MiniMax undercuts Kimi on price ($0.30 vs. $0.60/1M input for flagship). Also worth comparing: Zhipu GLM and Llama as open-weight alternatives at different capability and cost points.

Is the paid API tier worth it?

At $0.15/1M input tokens (M2.5), MiniMax is approximately 16x cheaper than Claude Opus 4.6 for input and roughly 8% of the total API cost per token according to independent analysis. For bulk processing, code review pipelines, or document analysis at scale, the economics are compelling.

The key caveat: MiniMax does not offer a genuinely free inference tier the way some open-source hosting services do. The Token Plan is a prepaid bulk package; the Unlimited Monthly Plan exists for heavy enterprise use. For evaluation and prototyping, OpenRouter hosts most MiniMax models with pay-as-you-go access starting at the per-token rates listed above, which is the lowest-friction way to test before committing to a direct platform account.

The MiniMax M2.7-highspeed variant is $0.60/1M input, double the standard M2.7 rate. Developers should confirm which variant is invoked by default when using third-party integrations, as the naming can be ambiguous in API wrapper documentation.

The open-weight Apache 2.0 models (M1, M2) can be self-hosted for teams with sufficient infrastructure, eliminating per-token costs entirely after deployment. This is the most cost-effective path for large-scale users but requires 230B-456B parameter deployment capacity.

Best use cases (and when to skip it)

Strong fit:

• Agentic software engineering pipelines, CI/CD automation, and multi-step code generation. TAU-bench performance supports this use case with verifiable data.

• Long-document analysis where a 200K-1M token context window eliminates the need for chunking. Entire codebases, legal documents, research corpora.

• Cost-sensitive production workloads where Claude or GPT-4 pricing is unsustainable and response quality on code tasks needs to remain near-frontier.

• Self-hosting or fine-tuning use cases where Apache 2.0 licensing is required.

• Teams already building on Anthropic SDK who want to swap in a cheaper model at the endpoint level without code changes.

Skip MiniMax when:

• Output throughput above 150 tokens/second is a hard requirement. DeepSeek V3.2 handles bursty high-speed inference better.

• The use case involves political, geopolitical, or Chinese-jurisdiction-sensitive topics where content filtering is likely to interfere.

• The primary task is mathematical proof generation or competition math. Kimi K2 Thinking and Qwen3.5 score measurably higher on AIME and formal reasoning benchmarks.

• Predictable rolling rate limits are a requirement. The fixed 5-hour window design requires rearchitecting retry logic for agent workflows.

• Local inference on consumer hardware is needed. The parameter counts are too large for all but enterprise GPU clusters.

Getting started with MiniMax

The quickest path: create an account at platform.minimax.io, purchase a Token Plan bundle, and call the API via the Anthropic-compatible endpoint (https://api.minimax.io/anthropic) using your existing Anthropic SDK code. Set the base URL and API key; the model string becomes "MiniMax-M2.7" or "MiniMax-M2.5" depending on your cost-performance target.

For evaluation without account creation, OpenRouter lists MiniMax M2 and M2.7 as pay-as-you-go models accessible with an OpenRouter key. This is the recommended first step before committing to a MiniMax platform account and Token Plan.

The open-weight M1 and M2 models are on HuggingFace under MiniMaxAI. For self-hosting, start with the M2 variant (230B total, 10B active) as a more tractable deployment target than M1's 456B. vLLM and SGLang have both been used successfully by the community for inference serving.

Rate limit planning: assume fixed 5-hour windows and build exponential backoff accordingly. The Coding Plan (prompt-quota-based) and direct API credit billing have different economics for different workload shapes; teams running sustained agent loops generally prefer credit billing to avoid plan-window blocking.