Coding agents
Use M2.7 for repository work, debugging, refactoring, tool calls, and other multi-step software-engineering tasks.
Independent tool overview
Minimax M2.7 at a glance
MiniMax M2.7 is an open-weight text model built for coding, tool use, and long-running agent workflows. It remains available at a low API price, although MiniMax M3 is now the newer, multimodal model with a much larger context window.
Visit the official Minimax M2.7 site ↗
- Best for
- Coding and text-based agents
- Context window
- 204,800 tokens total
- Model size
- 229B parameters
- Access
- API, MiniMax Agent, or open weights
- Current generation
- Succeeded by MiniMax M3
Overview
What Minimax M2.7 is
MiniMax released M2.7 in March 2026 as a 229-billion-parameter open-weight model focused on software engineering, office work, tool use, and multi-agent workflows. The hosted API supports a 204,800-token combined input-and-output window, while MiniMax publishes model weights for teams that want to run it on their own infrastructure.
MiniMax calls M2.7 its first model to participate deeply in its own evolution. That does not mean the public model autonomously retrains itself. MiniMax describes internal experiments in which M2.7 built and revised agent harnesses, skills, memory systems, evaluation loops, and research workflows while human researchers retained control over goals and critical decisions.
M2.7 is no longer MiniMax's newest M-series model. MiniMax M3 adds native image and video understanding, computer use, and a 1-million-token context window. M2.7 still makes sense when a text-only coding or agent workload values its established integrations and lower-cost standard or high-speed API options.
Use cases
Who Minimax M2.7 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Agent orchestration
Its focus on skills, tools, memory, and role consistency suits teams building text-based autonomous or multi-agent workflows.
Cost-conscious API workloads
The standard hosted model offers low per-token pricing for applications that do not need M3's multimodal input or 1M-token context.
Self-hosted experimentation
Published weights and deployment guidance support teams with enough infrastructure to run and evaluate the model themselves.
Capabilities
Core Minimax M2.7 features
Agent-oriented reasoning
M2.7 is tuned for multi-step work involving tools, skills, memory, search, and complex execution environments.
Software engineering
MiniMax reports gains across debugging, security analysis, machine learning, mobile development, repository-level generation, and terminal tasks. Treat its benchmark comparisons as vendor-reported results.
Office deliverables
The model can work through document, spreadsheet, and presentation tasks when connected to an agent harness with the relevant files and tools.
Native Agent Teams behavior
MiniMax trained the model for role adherence and collaboration in multi-agent setups, though reliability still depends on the surrounding orchestration and evaluation.
Compatible APIs
The hosted service accepts Anthropic-style and OpenAI-style requests; MiniMax recommends its Anthropic-compatible endpoint for thinking blocks and interleaved reasoning.
Standard and high-speed variants
The standard endpoint is documented at roughly 60 output tokens per second and the high-speed version at roughly 100 tokens per second.
Open-weight deployment
The official 229B-parameter checkpoint is available through MiniMax's Hugging Face organization, with guidance for SGLang, vLLM, Transformers, ModelScope, and NVIDIA NIM.
Process
How the Minimax M2.7 workflow works
Step 1
Choose hosted or self-hosted access
Use MiniMax's API for the quickest setup, or assess the hardware and operational cost of serving the 229B-parameter checkpoint yourself.
Step 2
Pick the endpoint
Select standard M2.7 for lowest cost or M2.7-highspeed when output latency matters more than token price.
Step 3
Connect tools deliberately
Give the model only the file, shell, browser, database, or MCP permissions required for the task, with approval gates for consequential actions.
Step 4
Evaluate on real work
Test repository tasks, tool-call accuracy, long-context behavior, latency, and cost on your own workload instead of relying only on vendor benchmarks.
Step 5
Compare with M3
Move to M3 when you need native image or video input, computer use, or a context window beyond 204,800 tokens.
Cost
Minimax M2.7 pricing and free plan
MiniMax sells M2.7 through pay-as-you-go API billing and separate subscription-style token plans. The prices below are the published pay-as-you-go API rates; self-hosting has no per-token vendor charge but requires substantial compute and operations.
MiniMax M2.7
$0.30 input / $1.20 output per 1M tokens
Standard hosted text model, documented at approximately 60 output tokens per second.
- Prompt-cache reads: $0.06 per 1M tokens
- Prompt-cache writes: $0.375 per 1M tokens
- 204,800-token combined context window
MiniMax M2.7 Highspeed
$0.60 input / $2.40 output per 1M tokens
Faster hosted variant with the same stated model performance and approximately 100 output tokens per second.
- Prompt-cache reads: $0.06 per 1M tokens
- Prompt-cache writes: $0.375 per 1M tokens
- 204,800-token combined context window
Open weights
No per-token MiniMax API fee
Download the official checkpoint and run it on compatible infrastructure.
- 229B-parameter model
- Compute, storage, networking, and engineering costs are your responsibility
- Review the model repository's license before commercial deployment
Pricing checked . Check current pricing at the source ↗
Assessment
Minimax M2.7 strengths and limitations
Where it stands out
- Low published API pricing for a coding- and agent-focused model
- Open weights plus hosted standard and high-speed endpoints
- Large 204,800-token combined context window
- Anthropic-compatible and OpenAI-compatible API access
- Strong emphasis on tool use, skills, memory, and multi-agent execution
What to consider
- MiniMax M3 is now the newer generation and adds multimodal input, computer use, and a 1M-token context window
- M2.7 is text-only and its 204,800-token limit counts input and output together
- The 229B-parameter checkpoint is expensive and operationally demanding to self-host
- Many performance claims and comparisons come from MiniMax's own evaluations
- The model does not autonomously retrain itself in a user's deployment; the self-evolution examples rely on surrounding harnesses, evaluations, and human direction
- Agents with shell, database, browser, or financial tools still require permission controls and human review
Compare
Minimax M2.7 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
MiniMax M3
Choose MiniMax's newer model for native multimodal input, computer use, and a 1M-token context window.
Explore MiniMax M3 →Consumer
GLM-5.3
Compare another current open-weight model built for coding, agents, long context, and multimodal tasks.
Explore GLM-5.3 →Consumer
Claude Opus 4.6
Consider Anthropic's closed model when premium coding and agent performance matters more than open weights or low token cost.
Explore Claude Opus 4.6 →Questions
Minimax M2.7 FAQs
What is MiniMax M2.7?
MiniMax M2.7 is a 229B-parameter open-weight text model designed for coding, tool use, office workflows, and multi-agent systems. It can be used through MiniMax's hosted API, MiniMax Agent, or self-hosted weights.
Is MiniMax M2.7 still available?
Yes. MiniMax's API documentation still lists M2.7 and M2.7-highspeed as available models. MiniMax M3 is now the latest M-series model, while M2.7 remains a lower-cost text-only option.
Is MiniMax M2.7 really self-evolving?
MiniMax uses that phrase for internal workflows where M2.7 helped build and improve agent harnesses, skills, memory, and evaluation loops. The public model does not independently retrain its weights while you use it, and human researchers still directed critical decisions in the examples MiniMax published.
How much does the MiniMax M2.7 API cost?
As of August 29, 2026, standard M2.7 costs $0.30 per million input tokens and $1.20 per million output tokens. M2.7-highspeed costs $0.60 input and $2.40 output per million tokens. Prompt caching has separate read and write rates.
What is the MiniMax M2.7 context window?
MiniMax documents a 204,800-token total context window for both the standard and high-speed variants. That maximum includes input and output tokens together.
Is MiniMax M2.7 open source?
MiniMax publishes the model weights and deployment instructions, so open-weight is the most precise description. Review the license attached to the official model repository before assuming unrestricted open-source or commercial rights.
Should I use MiniMax M2.7 or M3?
Use M2.7 when low-cost text generation, coding, or tool use is the priority. Use M3 when you need image or video understanding, desktop computer use, a 1M-token context window, or MiniMax's newest model generation.
Bottom line
Our Minimax M2.7 verdict
MiniMax M2.7 remains an appealing low-cost model for text-based coding and agent systems, especially for teams that value open weights and familiar API formats. Its 'self-evolution' story is best understood as agent-harness optimization under human direction, not autonomous retraining. New projects should compare it directly with M3, because M3 is now the current MiniMax generation and adds capabilities M2.7 cannot provide.
Visit Minimax M2.7 website ↗Similar Consumer Tools
View all Consumer AI tools →
GLM-5-Turbo
GLM-5-Turbo - Z AI's high-speed agentic model built specifically for OpenClaw
Cohere Transcribe
Cohere Transcribe- SOTA, free open-source speech recognition model
GPT-5.4 mini & nano
GPT-5.4 mini & nano - OpenAI's fast, cheap small models built for coding agents and subagent workflows
Qwen3.5-Omni
Qwen 3.5 Omni - Alibaba's native omnimodal model with text, image, audio, and video understanding across 113 languages
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.