Coding-agent experiments
Teams can pair M2.5 with a sandbox and tool harness to test repository analysis, code changes, and multi-step debugging workflows.
Independent tool overview
M2.5 at a glance
MiniMax M2.5 is a prior-generation but still available open-weight language model built for coding, tool use, search, and document-heavy agent workflows.
Visit the official M2.5 site ↗
- Developer
- MiniMax
- Released
- February 2026
- Current position
- Still available; listed as a legacy API model
- Model size
- About 230B total parameters, roughly 10B active per token
- API context
- 204,800 input-plus-output tokens
- License
- Modified MIT model license
Overview
What M2.5 is
MiniMax M2.5 is a reasoning-focused text model designed for coding and multi-step agent work. It can generate and refactor code, call tools, search through an agent harness, and help create or manipulate office documents when the surrounding application gives it the required tools. MiniMax distributes the model weights and also serves M2.5 through its API in standard and high-speed variants.
M2.5 is no longer MiniMax's newest model: the API now places it in the legacy-model section behind M2.7 and M3. It remains useful for teams that already depend on it, want reproducible open-weight deployment, or need a relatively inexpensive model for agent experiments. New projects should benchmark it against MiniMax's current models and other open-weight coding models before standardizing on it.
Use cases
Who M2.5 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Cost-sensitive API workloads
The standard API has low token prices for an agent-oriented reasoning model, especially when prompts can reuse cached context.
Open-weight deployment
Organizations that need control over hosting can download the weights and serve them with supported inference frameworks.
Long-context tool use
The hosted model supports a large shared input-and-output context for workflows that need substantial code, logs, or documents.
Existing M2.5 integrations
Teams already using its API behavior can keep a stable model target while evaluating migration to newer MiniMax releases.
Capabilities
Core M2.5 features
Agentic coding
M2.5 is trained for planning, code generation, refactoring, testing, and multi-step software tasks across several programming languages.
Function calling
The MiniMax text API accepts tool definitions and supports model-selected function calls for agent workflows.
Interleaved reasoning
The model can reason between tool calls, letting an agent inspect results and adjust its next action.
Large API context
MiniMax documents a 204,800-token combined context window for the hosted M2.5 and M2.5-highspeed endpoints.
Standard and high-speed endpoints
The two hosted variants use the same model capability, with the high-speed endpoint trading higher token prices for faster generation.
Prompt caching
The API offers separate cache-read and cache-write pricing, which can reduce repeated-context costs when implemented carefully.
Open model weights
MiniMax publishes M2.5 weights on Hugging Face under a modified MIT model license for local or private deployment.
Multiple serving frameworks
Official deployment guidance covers SGLang, vLLM, Transformers, KTransformers, and ModelScope.
Coding and office-work training focus
MiniMax positions the model for software work, search, and agent-assisted Word, spreadsheet, and presentation tasks when connected to suitable tools.
Process
How the M2.5 workflow works
Step 1
Choose hosted or self-hosted access
Use the MiniMax API for fast adoption or download the weights when infrastructure control matters more than setup simplicity.
Step 2
Define a bounded task
Give the model a clear objective, acceptance criteria, relevant context, and explicit limits on files, systems, or actions.
Step 3
Connect only necessary tools
Expose narrowly scoped functions for code, search, files, or business systems instead of unrestricted credentials and shell access.
Step 4
Run in an isolated environment
Execute generated code and agent actions in a sandbox with network, secret, and write controls appropriate to the risk.
Step 5
Verify the result
Review diffs, run tests, validate cited sources, and compare the final artifact with the original requirements before accepting it.
Step 6
Benchmark a newer model
Measure quality, latency, token use, and failure rate against M2.7, M3, or another candidate before committing to M2.5 for new production work.
Cost
M2.5 pricing and free plan
M2.5 remains available through MiniMax's pay-as-you-go API and as downloadable weights. The hosted model is now priced and labeled as a legacy option; infrastructure and engineering costs apply to self-hosting.
MiniMax-M2.5 API
$0.30 input / $1.20 output per 1M tokens
The standard hosted endpoint, documented at approximately 60 output tokens per second.
- $0.03 per 1M cache-read tokens
- $0.375 per 1M cache-write tokens
- Usage and account rate limits apply
MiniMax-M2.5-highspeed API
$0.60 input / $2.40 output per 1M tokens
A faster hosted variant with the same stated model performance and roughly 100 output tokens per second.
- $0.03 per 1M cache-read tokens
- $0.375 per 1M cache-write tokens
- Costs twice the standard endpoint for uncached input and output
Self-hosted weights
No per-token model fee
Downloadable model weights for teams prepared to operate large-model inference infrastructure.
- Modified MIT model license applies
- Compute, storage, networking, monitoring, and engineering are not free
- Review the license before commercial deployment
Pricing checked . Check current pricing at the source ↗
Assessment
M2.5 strengths and limitations
Where it stands out
- Still available through both the official API and downloadable model weights
- Built specifically around coding, tool calling, search, and multi-step agent work
- Large hosted context window for codebases, logs, and document-heavy tasks
- Low published token prices compared with many frontier hosted models
- Standard and faster API variants make latency-versus-cost tradeoffs explicit
- Prompt caching can lower the cost of repeated long context
- Official deployment guidance covers several widely used inference frameworks
- Open weights allow private deployment and independent evaluation
What to consider
- M2.5 is a prior-generation model and appears in MiniMax's legacy API pricing section
- MiniMax's benchmark figures are vendor-reported and do not guarantee performance on a specific codebase or workflow
- Long context capacity does not guarantee accurate retrieval or reasoning across every supplied token
- The roughly 230B-parameter checkpoint requires substantial storage and accelerator memory despite its sparse activation
- The model license is modified MIT, so teams should review its additional terms rather than assume standard MIT conditions
- Agent reliability depends heavily on the surrounding prompt, tool definitions, sandbox, retry logic, and verification layer
- Generated code can introduce security bugs, licensing issues, or regressions and still requires review and testing
- Tool-connected models can be vulnerable to prompt injection and unsafe instructions in retrieved content
- The hosted API requires sending prompts and outputs to MiniMax, which may not fit every privacy, residency, or procurement policy
- API pricing excludes downstream tool charges, search fees, storage, and the operational cost of failed or repeated agent runs
- Office-file capabilities require an application or agent harness that can actually read, edit, and validate those formats
- Teams starting fresh may get better results from MiniMax M2.7, M3, or another newer model
Compare
M2.5 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
Minimax M2.7
MiniMax's newer M2-series model is the most direct migration path for teams that want updated agent and coding performance.
Explore Minimax M2.7 →Coding
Qwen3.6-27B
A much smaller open-weight coding model to evaluate when deployment footprint and local serving practicality matter.
Explore Qwen3.6-27B →Agents
Kimi K2.5
Another open agent-focused model family for teams comparing tool use, coding quality, and hosting options.
Explore Kimi K2.5 →Business Operations
DeepSeek
A broad open-model ecosystem with chat and API access for developers comparing cost and reasoning capability.
Explore DeepSeek →Questions
M2.5 FAQs
What is MiniMax M2.5?
MiniMax M2.5 is an open-weight text model released in February 2026 for coding, tool use, search, reasoning, and agent-assisted office work. It is available through MiniMax's API and as downloadable weights.
Is MiniMax M2.5 still available?
Yes. MiniMax still lists M2.5 and M2.5-highspeed as supported text API models and publishes the weights. However, its pricing page now classifies M2.5 as a legacy model behind newer releases.
Is MiniMax M2.5 open source?
The model weights and deployment materials are publicly downloadable. MiniMax labels the model license Modified MIT, so it is better described as open-weight and users should read the exact license terms before deployment.
How much does the MiniMax M2.5 API cost?
As of August 30, 2026, the standard endpoint costs $0.30 per million input tokens and $1.20 per million output tokens. The high-speed endpoint costs $0.60 input and $2.40 output per million tokens, with separate cache pricing.
What context window does M2.5 support?
MiniMax's current API overview lists a 204,800-token combined input-and-output context window for both M2.5 endpoints. Self-hosted limits can depend on the serving configuration and available memory.
Can M2.5 run locally?
Yes, but it is not a typical laptop-scale model. The roughly 230B-parameter checkpoint requires substantial storage and accelerator resources; official guidance focuses on frameworks such as SGLang and vLLM.
Is M2.5 good for autonomous coding?
It is designed for coding agents, but no language model should receive unrestricted production access. Use a sandbox, narrow permissions, automated tests, human review, and a rollback path.
Should a new project use M2.5 or a newer model?
Treat M2.5 as a strong legacy candidate, not the automatic default. Run representative evaluations against M2.7, M3, and other current models using the same harness, tools, prompts, and cost accounting.
Bottom line
Our M2.5 verdict
MiniMax M2.5 remains a credible low-cost, open-weight model for coding and agent research, especially for existing integrations or teams that need deployment control. Because MiniMax now treats it as a legacy API model, new production projects should use it as a benchmark candidate and choose only after testing newer alternatives on real tasks.
Visit M2.5 website ↗Similar Coding Tools
View all Coding AI tools →
Orchids 1.0
Orchids 1.0 - AI app builder for any stack with BYO model subscriptions
GPT-5.3-Codex-Spark
GPT 5.3 Codex Spark - OpenAI’s ultra-fast model for real-time coding
Oz
Oz - Warp's cloud platform for orchestrating hundreds of coding agents in parallel
Rork Max
Rork Max - Rork's AI-powered native iOS app builder for every Apple platform
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.