MiniMax M2.7 suits teams who want frontier-adjacent agentic coding without a frontier API bill: swap it into a coding-agent setup like Cline or OpenCode when budget matters more than raw speed. Its GDPval-AA Elo of 1495 signals solid real-world task competence, but expect slower, more verbose output than a hosted flagship model.
MiniMax M2.7 is a Sparse Mixture-of-Experts language model built specifically for agentic coding and tool use rather than general multimodal chat. On PinchBench it posts 86.2%, just 1.2 points shy of Claude Opus 4.6's result, and ships as open weights under a Modified-MIT license instead of staying closed API-only like most frontier-adjacent releases.
Where it sits
- $0.525/M$ per 1M tokensBlended price (3:1)Lower is better#15 / 64peer median $1.70/Mvendor price, checked by HokAI
- 46 tok/stokens/sOutput speedHigher is better#33 / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 44peer median 88.3%per source, see benchmark scores
Cheaper than 75% of the 64 GA models with a published price, rank 33 of 39 on output speed as cited from Artificial Analysis, and one of 65 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: MiniMax · Family: MiniMax M2
Context window: 204,800 tokens · Max output: 196,608
Input modalities: text, code, tool-calls · Output: text, code, tool-calls
About MiniMax M2.7
MiniMax-M2.7 is a large language model built by MiniMax, a Shanghai-based AI company, and released on March 18, 2026. It is a Sparse Mixture-of-Experts model with 230 billion total parameters but only 10 billion active per token, spread across 256 experts with 8 activated per pass, using 62 layers and a hidden size of 3072. The architecture combines multi-head causal self-attention with Rotary Position Embeddings (RoPE) and Query-Key RMSNorm, plus top-k expert routing. M2.7 is part of MiniMax's M-series, following M2, M2.1, and M2.5, and preceding the multimodal M3 released June 1, 2026. MiniMax dedicated all of M2.7's training compute to 'Code + Agent' domains, positioning it as a coding and agentic-workflow specialist rather than a general multimodal flagship.
On benchmarks, M2.7 scores 56.2% on SWE-Bench Pro, a 23.6-point jump over M2.1's 32.6%, and 57.0% on Terminal-Bench 2, up from M2.1's 47.9%. It reaches 86.2% on PinchBench, within 1.2 points of Claude Opus 4.6, and posts an Artificial Analysis Intelligence Index of 50, eight points above M2.5, putting it ahead of MiMo-V2-Pro (49) and Kimi K2.5 (47) and roughly level with GLM-5 (50). MiniMax also reports a GDPval-AA Elo of 1495. Independently reported scores on graduate-level reasoning benchmarks such as GPQA Diamond are notably lower than its coding scores, reflecting the training compute split toward code and agents.
The model has a 204,800-token context window and uses full attention across it, so latency rises as context fills; reviewers note that pushing near that limit feels slow in practice. M2.7 is text-only: it accepts plain text and code, with no native image, audio, or video input or output. Tool use and function calling are fully supported via the documented tool_calling_guide.md schema, using the same syntax as MiniMax M2, and the model is built for agent harnesses including Cline, OpenCode, and Kilo. For image understanding, MiniMax documents a workaround of registering an 'analyze_image' function-call tool that proxies to a vision model such as Claude, GPT, or Gemini.
M2.7 was initially released as a proprietary API model, then published with open weights on Hugging Face and ModelScope in April 2026. The full BF16 release is roughly 457GB, while community GGUF quantizations range down to about 60GB for a 1-bit build, making it runnable on prosumer hardware via vLLM or SGLang (NVIDIA also ships an NVFP4 quantization). Shortly after the open-weight release, MiniMax updated the Hugging Face license from MIT to a 'Modified-MIT' license requiring the company's written authorization for commercial use, which drew criticism from developers who had begun building under the original terms.
MiniMax states that customer API inputs are not stored or used for training unless a customer opts in, and the company offers data residency choices across North America, Europe, and Asia-Pacific. No SOC 2 Type II report, ISO 27001 certificate, or HIPAA-eligible tier was found in public sources, and MiniMax hasn't released a public system-card document or named outside red-team partners for M2.7. As a Shanghai-based company, MiniMax also operates under China's Generative AI Measures from the Cyberspace Administration of China.
M2.7 is best for budget-conscious teams building agentic coding pipelines, developers who want to self-host an open-weight frontier-adjacent model, and teams already using Cline, OpenCode, or Kilo. It is a poor fit for interactive low-latency chat, any task requiring image, audio, or video understanding, and workloads needing top-tier graduate-level reasoning benchmarks like GPQA Diamond or AIME.
MiniMax's M-series cadence has been rapid: M2.5 in February 2026, M2.7 in March 2026 (open-weighted in April), and M3 on June 1, 2026 with a new MiniMax Sparse Attention (MSA) architecture, a 1M-token context window, and native multimodality. M2.7 remains available and has not been deprecated, but M3 is now MiniMax's flagship for multimodal and longer-context workloads, while M2.7 continues to serve as the company's cost-efficient, text-first agentic coding option.
Pricing
MiniMax's direct API charges $0.30 per 1M input tokens and $1.20 per 1M output tokens on the standard tier, with cached input reads at $0.06 per 1M tokens and cache writes at $0.375 per 1M tokens. A lower-latency 'HighSpeed' variant runs about $0.60 input and $2.40 output per 1M tokens. Self-hosting the open weights is free, but commercial deployments need MiniMax's written go-ahead under the Modified-MIT license.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0090 | $0.0012 | $0.010 |
| Support reply | $0.0006 | $0.0004 | $0.0010 |
| One coding agent run | $0.060 | $0.024 | $0.084 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Self-Evolving Training Loop: M2.7 is MiniMax's first model that actively participates in its own development, handling an estimated 30-50% of its own reinforcement learning research workflow.
- Full-Attention Long Context: One of the larger context windows in MiniMax's M-series, sized for large multi-file codebases and long agent transcripts, with a 196,608-token max output per response.
- Sparse MoE Routing: Routes each token through 8 of 256 experts across 62 layers, keeping inference cost down while preserving the model's total capacity.
- Open Weights on Hugging Face and ModelScope: Ships under a Modified-MIT license, with community builds ranging from a compact 1-bit GGUF up to the full-precision release, runnable via vLLM or SGLang.
- Tool Calling for Agent Harnesses: Supports the documented MiniMax-M2 tool-call syntax for structured function calling, used by agent frameworks like Cline, OpenCode, and Kilo.
Pros
- Costs a fraction of what Claude Opus or GPT-5 charge per token for comparable agentic coding tasks, per MiniMax's own pricing page.
- Terminal-Bench 2 score of 57.0 (up from M2.1's 47.9) shows real gains on agentic terminal tasks, not just isolated benchmarks.
- Open weights mean you can self-host entirely on your own hardware and avoid API lock-in, unlike most frontier-adjacent coding models.
- A 196,608-token max output leaves room for long multi-file diffs and full agent transcripts in a single response, not just a large prompt window.
Cons
- Output speed of 35-46 tokens/sec with roughly 2.2s time-to-first-token is sluggish next to similarly priced open-weight peers.
- Text-only: no native image, audio, or video input, unlike the newer MiniMax M3.
- Highly verbose output, generating up to 4x the average tokens of similarly-sized models on benchmark evals, which raises real-world cost and latency.
- April 2026 license change to 'Modified-MIT' requires written authorization from MiniMax for commercial self-hosted deployments.
Benchmarks
- Pinchbench: 86.2 vendor-reported · 12 Apr 2026
- Gdpval Aa Elo: 1495 vendor-reported · 12 Apr 2026
- SWE-bench Pro: 56.2% vendor-reported · 12 Apr 2026 — Harder, longer real-repository coding tasks, % solved.
- Terminal Bench 2: 57 vendor-reported · 12 Apr 2026
- AA Intelligence Index: 50 cited: Artificial Analysis · 13 Jun 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- Output speed: 46 tok/s cited: Artificial Analysis · 13 Jun 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What are MiniMax M2.7's pricing plans in 2026?
MiniMax charges by the token on M2.7's API, with a faster low-latency HighSpeed variant priced higher for interactive use. A single agentic coding session of roughly 500,000 input tokens and 100,000 output tokens runs about $0.27 on the standard tier, and third-party hosts like Fireworks or Together.ai blend to around $0.22 per 1M tokens. You can also self-host the open weights for free, though the Modified-MIT license requires MiniMax's sign-off for commercial deployments.
Is MiniMax M2.7 free to use?
MiniMax doesn't offer a free hosted tier for M2.7 on its own API. The model does ship as open weights you can self-host at no cost, with community quantized builds running from about 60GB up to the full 457GB BF16 release, though commercial self-hosted use needs MiniMax's written authorization under the Modified-MIT license.
What are the best alternatives to MiniMax M2.7?
DeepSeek V4 is a fully MIT-licensed open-source alternative with a higher published coding-benchmark score. MiniMax M3 adds native multimodal input on the same team's architecture and a much larger context window. GLM-5.2 offers a 1M-token context window under an open MIT license for teams that need to hold more of a codebase in context at once.
MiniMax M2.7 or DeepSeek V4: which should you pick?
DeepSeek V4 posts a higher 80.6% SWE-bench Verified score than M2.7's SWE-Bench Pro result and ships under a plain MIT license with no commercial-use clause, while M2.7 counters with a lower per-input-token price and an established tool-calling track record in Cline and OpenCode. Pick DeepSeek V4 for the stronger open benchmark score and unrestricted license; pick MiniMax M2.7 if you're already building on its documented tool-calling schema or want the cheaper input rate.
What does it take to start using MiniMax M2.7?
For the hosted API, sign up on MiniMax's platform, generate a key, and follow the tool_calling_guide.md schema to wire it into an agent framework like Cline, OpenCode, or Kilo. To self-host, pull the weights from Hugging Face or ModelScope, pick a GGUF quantization sized to your GPU, and serve it with vLLM or SGLang.
Top Alternatives
- DeepSeek V4: Pick DeepSeek V4 for a higher published SWE-bench Verified score (80.6%) and a fully permissive open license that skips MiniMax's authorization requirement; pick MiniMax M2.7 if you're already standardized on its tool-calling schema in Cline or OpenCode.
- MiniMax M3: Pick MiniMax M3 if you need native image input and a larger context window; pick MiniMax M2.7 if you only need text and want the model with the established Cline and OpenCode integration history.
- Claude Opus 4.6: Pick Claude Opus 4.6 if you need its higher 80.8% SWE-bench Verified score and don't mind double-digit per-output-token pricing; pick MiniMax M2.7 for a small fraction of that cost on comparable agentic coding tasks.
- GLM-5.2: Pick GLM-5.2 if you need a 1M-token context window under a fully open-source MIT license; pick MiniMax M2.7 for a lower per-input-token rate on agentic coding workloads.