MiniMax Agent
Type: full-code · Vendor: MiniMax · Language: API + Web product · License: proprietary (product) / open-weight (MiniMax-M1, M2) · Status: active · Status in practice: emerging
Shanghai-based MiniMax bundles a foundation-model line (ABAB → MiniMax-01 → M1 → M2.x) with a consumer MiniMax Agent that evaluates tasks, assembles an 'Agent Team' for them, and learns user-specific habits into custom skills.
Description. MiniMax was founded in 2021 by SenseTime alumni and listed in Hong Kong in January 2026. It develops a series of multimodal foundation models with strong code and Agent capabilities; the company markets MiniMax-M1 as the world's first open-source, large-scale, hybrid-attention reasoning model with an 80K-token reasoning output and tool-use benchmark leadership among open-weights. MiniMax Agent is the consumer surface: an assistant that evaluates a task, builds an agent team, learns user habits, and turns repetitive workflows into custom skills. The API exposes function calling for tool use; verifiable English platform docs were patchy at access time (several pages 404 or redirect repeatedly between intl.minimaxi.com / platform.minimax.io).
Agent loop shape. Two surfaces. (1) MiniMax Agent product: hosted loop that evaluates the user task, assembles an Agent Team, and dispatches sub-tasks; surfaces skills, memory and schedules in a single chat. (2) MiniMax API (M1 / M2.x via Chat Completions): standard tool-call loop; reasoning models expose a long 80K-token thinking budget and benchmark in TAU-bench style agentic tool-use scenarios.
Primary use cases
- consumer multi-skill assistant (work and life tasks)
- agent-team task decomposition over arbitrary user goals
- open-weight reasoning model for self-hosted tool-use agents
- long-output reasoning (80K token thinking budget)
Key concepts
- MiniMax-M1 reasoning model → extended-thinking (docs) — First open-source large-scale hybrid-attention reasoning model with 80K reasoning tokens and strong agentic tool-use benchmarks.
- MiniMax-M2.x line (docs) — Newer self-improvement-oriented reasoning + agent model line announced for production agent harnesses.
- Agent Team → orchestrator-workers (docs) — Product feature where MiniMax Agent evaluates a task and assembles a team of agents to solve it.
- Custom Skills → skill-library (docs) — Repetitive workflows the agent learns from the user's habits and turns into reusable skills.
- Function Calling → tool-use (docs) — Tool-use API on the MiniMax platform; documented under the Function Calling endpoint reference.
Patterns this full-code implements —
- ★★Structured Output
Function calling on MiniMax-M2.7 returns tool_calls with JSON-format arguments; the platform documents the schema of tools, tool_calls and function.arguments fields directly.
- ★★Tool Use
Documented as a top-level API capability ('agent tool use scenarios' in M1 release) and as a benchmarked competence of the open-weight models.
- ★★Orchestrator-Workers
MiniMax Agent's headline product behaviour is 'Build Agent Team: Evaluate tasks and assemble teams to solve problems' — an orchestrator-over-workers shape. The M2 launch blog also documents coordinat…
- ★Skill Library
'Know Your Habits' turns repetitive tasks into custom skills accessible from the same chat surface.
- ★★Extended Thinking
MiniMax-M1 surfaces an 80K-token reasoning output budget, claimed to be an industry leader for reasoning length.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified