LLMPrivate

MiniMax M2.5

MiniMax M2.5 is a high-performance, agent-native language model optimized for coding, tool use, and real-world productivity tasks with SOTA scores in agentic benchmarks.

Maker
MiniMax AI
Modality
Text
Context
198K tokens
Open weights
Yes — Modified-MIT

Overview

What is MiniMax M2.5

MiniMax M2.5 is a frontier AI model from MiniMax AI, released in February 2026. It excels in coding, agentic tool use, and complex task decomposition, achieving state-of-the-art results in benchmarks like SWE-Bench Verified (80.2%) and BrowseComp (76.3% with context management), with efficient reasoning and low operational cost.

Running it privately on Venice

On Venice, MiniMax M2.5 runs with full privacy—zero retention of your prompts or outputs. You get uncensored access to its full capabilities, including tool calling, web search, and code optimization, without surveillance or data harvesting. This is AI sovereignty in action: powerful reasoning, private by design.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • State-of-the-art performance in coding benchmarks: 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench.
  • Highly efficient agentic reasoning: completes the SWE-Bench Verified evaluation 37% faster than M2.1 (22.8 vs 31.3 minutes, per MiniMax).
  • Strong multilingual coding support across Go, C, C++, TypeScript, Rust, Kotlin, and more.
  • Spec-writing tendency enables architectural planning before code generation, improving project structure.
  • Optimized for real-world workflows with function calling, web search, and code execution capabilities.
Limitations
  • High measured hallucination rate: 88% on Artificial Analysis's AA-Omniscience benchmark (third-party; up from M2.1's 67%) — so verify outputs in knowledge-critical tasks.
  • Smaller context window (198K) compared to rivals like Claude models (1M) and DeepSeek V4 Flash (1M).
  • Modified-MIT license carries usage conditions beyond plain MIT — review it before commercial self-hosting.
  • Limited vision or audio input support—text-only modality.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
MiniMax AI
Released
February 2026
Architecture
Mixture-of-Experts (MoE)
Parameters
229B total / 10B active
Open weights
Yes — Modified-MIT
Context window
198K tokens
Max output
32.768K tokens
Capabilities
Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Feb 2026
License
Modified-MIT

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-m25",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.27 per 1M input tokens and $0.95 per 1M output tokens.

Input / 1M tokens
$0.27
Per 1M tokens
Output / 1M tokens
$0.95
Per 1M tokens
Cached input / 1M
$0.03
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelContext windowStrongest atOpen weightsPrice (Venice)
MiniMax M2.5198K tokensAgentic coding & tool useYes$0.27 in · $0.95 out / 1M
Claude Opus 51M tokensLong-context reasoningNo$6 in · $30 out / 1M
DeepSeek V4 Flash 07311M tokensSpeed & cost efficiencyNo$0.17 in · $0.35 out / 1M
Google Gemma 4 31B Instruct256K tokensOpen-weight codingYes$0.12 in · $0.36 out / 1M

SOTA in coding and agentic workflows with efficient reasoning and low cost per task.

Use cases

What it is good for

  1. 01Automated software development and bug fixing via SWE-Bench-style workflows.
  2. 02Agentic coding assistants that plan, write, and refactor code across multiple files.
  3. 03Real-time web research and data synthesis using integrated search tools.
  4. 04Internal office automation for document drafting, analysis, and reporting.
  5. 05High-throughput API workloads where cost efficiency and speed are critical.

Prompting

Getting better results

Use explicit task decomposition—ask the model to plan first, then implement.

Leverage its architect-like reasoning by requesting spec drafts before coding.

For multilingual tasks, specify the programming language clearly in the prompt.

Use function calling to chain tools and automate complex workflows step-by-step.

Version history

MiniMax M2.1
2025

Predecessor with lower agentic performance and hallucination rate.

MiniMax M2.5
2026-02

Current version — optimized for speed, coding, and real-world workflows.

FAQ

Frequently asked questions

MiniMax M2.5 is a large language model developed by MiniMax AI, released in February 2026. It is optimized for coding, agentic workflows, and real-world productivity tasks, with state-of-the-art performance in benchmarks like SWE-Bench and BrowseComp.

On Venice, MiniMax M2.5 costs $0.27 per million input tokens and $0.95 per million output tokens. Cached input is even cheaper at $0.03 per million tokens.

MiniMax M2.5 is open-weights: the model is published on Hugging Face under a Modified-MIT license and can be self-hosted. On Venice you can use it without running your own infrastructure, at $0.27 per 1M input tokens and $0.95 per 1M output tokens.

Yes, MiniMax M2.5 supports tool use and function calling, making it ideal for building agentic systems that interact with code, APIs, and web search.

MiniMax M2.5 has a context window of 198,000 tokens, allowing for long conversations and document processing, though smaller than some rivals with 1M-token windows.

Yes—MiniMax M2.5 runs on Venice with zero retention. Your prompts are never stored, profiled, or used for training, ensuring full privacy and sovereignty.

MiniMax M2.5 is faster and far cheaper for coding and agentic tasks, while Claude Opus 5 offers deeper reasoning and a 1M-token context. Choose M2.5 for cost-efficient automation, Opus for complex analysis.

MiniMax M2.5 supports multiple programming languages including Go, C, C++, TypeScript, Rust, Kotlin, and others, with strong multilingual performance in coding benchmarks.

Run MiniMax M2.5 privately

No prompt logging. No data used for training.