LLMPrivate

MiniMax M2.7

MiniMax M2.7 is a self-evolving, code-optimized reasoning model with strong agentic capabilities, delivering near-opus-level performance at a fraction of the cost.

Maker
MiniMax
Modality
Text
Context
198K tokens
Open weights
No

Overview

What is MiniMax M2.7

MiniMax M2.7 is a powerful reasoning and code-optimized text model developed by MiniMax, released in March 2026. It is designed for complex productivity tasks, capable of self-evolution, agent team orchestration, and real-world software engineering, with a 198K context window and full tool use support.

Running it privately on Venice

On Venice, MiniMax M2.7 runs with full privacy — your prompts are never stored, profiled, or used for training. You get uncensored access to its agentic and code-optimized capabilities, including web search and function calling, without sacrificing sovereignty. The model’s zero retention on Venice ensures your inputs remain private, even during long-running agent workflows.

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

Assessment

Strengths and limitations

Strengths
  • Exceptional real-world software engineering: strong on SWE-Pro (56.22%), VIBE-Pro (55.6%), and system-level reasoning for SRE tasks.
  • Agentic self-evolution: capable of updating its own memory, building skills, and improving its learning process.
  • High skill adherence rate (97%) in complex tool-use scenarios, making it reliable for agent workflows.
  • Cost-efficient performance: delivers near-Claude Opus-level coding quality at ~1/20th the cost per token.
  • Supports function calling, web search, and code-optimized reasoning for end-to-end task automation.
Limitations
  • Not open-source: weights are proprietary and commercial use requires written authorization from MiniMax.
  • Smaller context window (198K) compared to rivals like DeepSeek V4 Flash (1M) and Claude models (1M).
  • Slightly lower benchmark scores than frontier models like Opus-4.6 and GPT-5.4 in reasoning and math.
  • No end-to-end encryption or TEE protection on Venice, limiting privacy to zero retention only.

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
Released
March 2026
Architecture
Mixture-of-Experts (MoE) — 230B total / 10B active
Parameters
230B (10B active)
Open weights
No
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
Mar 2026
License
Modified MIT (commercial use requires authorization)

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-m27",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.38 per 1M input tokens and $1.50 per 1M output tokens.

Input / 1M tokens
$0.38
Per 1M tokens
Output / 1M tokens
$1.50
Per 1M tokens
Cached input / 1M
$0.07
Per 1M tokens

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

Alternatives

How it compares

ModelMax resolutionStrongest atOpen weightsPrice (Venice)
MiniMax M2.7N/AAgentic coding, self-evolutionNo$0.38 in · $1.50 out / 1M
DeepSeek V4 Flash 0731N/ARaw coding speed, mathNo$0.17 in · $0.35 out / 1M
Claude Opus 5N/AComplex reasoning, accuracyNo$6 in · $30 out / 1M
GLM 5.1N/AOpen weights, multilingualYes$1.10 in · $4.15 out / 1M

Self-improving agent model with strong code and reasoning at ultra-low cost.

Use cases

What it is good for

  1. 01Automating complex software engineering tasks like bug triage, refactoring, and security reviews.
  2. 02Building autonomous agent teams that use dynamic tool search and multi-step reasoning.
  3. 03Reducing incident recovery time in production systems with real-time log and trace analysis.
  4. 04Running private, uncensored coding agents on Venice without data retention risks.
  5. 05High-fidelity office suite automation involving Excel, PPT, and Word with multi-turn edits.

Prompting

Getting better results

Use explicit agent team patterns — define roles and handoffs in your prompt for complex tasks.

Leverage web search and tool use by asking M2.7 to retrieve current data or run code.

For coding tasks, reference SWE-Pro or VIBE-Pro style workflows to align with its training.

Request step-by-step reasoning to improve accuracy and reduce hallucinations in critical tasks.

Version history

MiniMax M2.5
2026-02

Predecessor model with lower reasoning and agentic performance.

MiniMax M2.7
2026-03

Current — self-evolving, agentic, code-optimized.

FAQ

Frequently asked questions

MiniMax M2.7 is a reasoning and code-optimized language model released by MiniMax in March 2026. It is designed for complex agent workflows, self-evolution, and real-world software engineering tasks, supporting tool use, web search, and dynamic skill orchestration.

On Venice, MiniMax M2.7 costs $0.38 per 1M input tokens and $1.50 per 1M output tokens, with cached input at $0.07 per 1M. Pricing is transparent and billed per token, with no subscription required.

No, MiniMax M2.7 is not open source. It uses a modified MIT license that allows non-commercial use, but commercial use requires prior written authorization from MiniMax. The model weights are not publicly available.

Yes, MiniMax M2.7 supports function calling, web search, and code execution, making it well-suited for building autonomous agents and completing multi-step productivity tasks.

MiniMax M2.7 has a context window of 198K tokens, allowing for long conversations and document processing, though smaller than some rivals like DeepSeek V4 Flash and Claude models.

Yes. On Venice, MiniMax M2.7 runs with zero retention — your prompts are never stored or used for training. This ensures full privacy while maintaining access to its full agentic and reasoning capabilities.

Claude Opus 5 leads in deep reasoning and accuracy on long-horizon tasks, but MiniMax M2.7 delivers nearly comparable performance at a fraction of the cost, especially in coding and agentic workflows. For budget-conscious, high-throughput automation, M2.7 is highly competitive.

MiniMax M2.7 scores 56.22% on SWE-Pro, 55.6% on VIBE-Pro, and achieves a 66.6% medal rate on MLE Bench Lite. It also reaches an ELO of 1495 on GDPval-AA, the highest among open models at its release.

Run MiniMax M2.7 privately

No prompt logging. No data used for training.