Streaming chat workspace
Test MiniMax-M2’s reasoning loop with live <think> traces, conversation search, and persona switching designed for red-teaming.
MiniMax-M2: Efficient Agent-Grade Intelligence
A 230B-parameter sparse MoE with 10B active parameters, tuned for coding, browsing, and long-horizon agents.
230B sparse MoE • 10B active per token • 128K interleaved context • SWE-bench Verified 69.4 · Terminal-Bench 46.3 · BrowseComp 44
Ship development copilots, research agents, and finance analysts with MiniMax-M2’s open weights, free MiniMax Agent console, and limited-time complimentary APIs.
Or launch a prebuilt MiniMax-M2 expert profile
Popular
Mathematics Tutor
Helps you solve complex maths problems with step-by-step reasoning and LaTeX formatting.
Popular
Code Expert
Generates production-ready code, reviews implementations, and explains trade-offs.
New
Frontend Designer
Builds polished UI components with responsive layouts and accessibility baked in.
Document Analyst
Extracts insights from long documents, summarises content, and supports multilingual analysis.
Creative Writer
Crafts marketing copy, scripts, and narratives tailored to your tone and audience.
MiniMax-M2 Assistant
A general-purpose MiniMax-M2 assistant ready to help with reasoning, coding, and knowledge tasks.
Try guided MiniMax-M2 prompts
Experience MiniMax-M2 in action
Hands-on demos using the same harnesses as our public benchmarks
Launch ready-made scenarios that highlight interleaved thinking, tool execution, and coding depth in realistic environments.
API quickstart
Replica endpoints for OpenAI and Anthropic contracts let you drop MiniMax-M2 into existing SDKs in minutes.
Agent notebooks
Run notebook-driven agent evaluations with BrowseComp and Terminal-Bench scaffolds to audit MiniMax-M2 locally.
Engineering for dependable agents
Architecture decisions behind MiniMax-M2
MiniMax-M2 combines a sparse 230B-parameter core with interleaved thinking traces so teams can debug, govern, and extend agents with confidence.
Sparse MoE core
1/32 expert routing ensures only ~10B parameters fire per token while the full 230B capacity remains available when needed.
- Maintains latency comparable to 50B dense models while retaining frontier reasoning quality.
- Supports long tool chains without blowing past GPU memory budgets.
- MoE telemetry exposes which experts activated for audit trails.
Benchmark-proven execution
Independent evaluations rank MiniMax-M2 #1 among open models for agentic workloads across SWE-bench, Terminal-Bench, and BrowseComp.
- 69.4 SWE-bench Verified and 36.2 Multi-SWE-Bench for multi-file repo fixes.
- 46.3 Terminal-Bench performance with claude-code harness over eight runs.
- 44 BrowseComp and 48.5 BrowseComp-zh for evidence-grounded browsing.
Interleaved thinking
Responses wrap reasoning in <think> tags so you can replay, redact, or condition future turns without losing context.
- Preserve <think> blocks in conversation history to maintain performance.
- Supports structured audits by exposing intermediate hypotheses.
- Fine-tune or post-process separately on thought vs. final answer segments.
Deployment-ready tooling
Day-zero support from vLLM, SGLang, and MLX lets you serve MiniMax-M2 anywhere—from cluster-scale inference to laptop experiments.
- Optimised kernels and KV-cache policies tuned for MiniMax-M2’s routing pattern.
- Reference configs for GPU, TPU, and Apple Silicon pipelines.
- Integration guides for managed APIs on the MiniMax Open Platform.
Deploy MiniMax-M2 with confidence
From instruction intake to audit-ready outputs
MiniMax-M2 keeps tool calls, citations, <think> traces, and final answers structured, which makes audits, safety reviews, and fine-tuning pipelines straightforward for autonomy teams.
Tool-aware, traceable responses
Public scores on BrowseComp, FinSearchComp, τ²-Bench, and HLE prove MiniMax-M2 can recover from flaky tools, cite sources, and keep multi-hop reasoning grounded.
What happens on every request
- 1Ingests instructions, retrieved documents, and prior <think> traces up to 128K tokens without truncation.
- 2Routes tokens across 32 experts so only ~10B parameters activate, maintaining low latency while preserving capacity.
- 3Returns structured reasoning, tool calls, and final answers optimised through interleaved thinking and sentence-level alignment.
Production playbooks
Risk & finance briefings
Transform filings, macro data, and internal metrics into investor-ready narratives with explicit assumptions and optional tool outputs for audit.
Review risk sampleTechnical architecture reviews
Turn ambiguous requirements into architecture memos, dependency maps, and prioritized remediation plans with linked reasoning traces.
Draft an architecture noteMultilingual policy analysis
Summarise and compare regional regulations, cite paragraphs inline, and highlight unresolved governance tasks for compliance teams.
Audit a policyWhere MiniMax-M2 excels
Benchmarks that translate into production workflows
MiniMax-M2 pairs trillion-parameter capacity with a 10B active set, delivering reliable plan-act-verify behaviour across coding, browsing, and analytics agents.
Autonomous repo repair
69.4% on SWE-bench Verified across OpenHands and R2E-Gym scaffolds proves MiniMax-M2 can complete multi-file run–fix loops inside real repositories.
Interactive terminals
A 46.3 Terminal-Bench score (8-run mean) shows MiniMax-M2 maintaining command accuracy under 100-step limits with flaky retries accounted for.
Research copilots
44 BrowseComp and 48.5 BrowseComp-zh demonstrate evidence-driven browsing in English and Chinese, with citations tracked for review.
Financial intelligence
65.5 on FinSearchComp-global confirms MiniMax-M2’s ability to surface signals across filings, earnings transcripts, and macro reports.
Multilingual engineering
56.5 SWE-bench Multilingual plus 36.2 Multi-SWE-Bench validate MiniMax-M2’s coverage for cross-language codebases and documentation.
Scientific analysis
75.7 GAIA (text-only) and 77.2 τ² Bench show MiniMax-M2 distilling technical papers with verifiable reasoning chains.
How teams ship with MiniMax-M2
From evaluation to production in three steps
Provision access, instrument usage, and keep budgets predictable with token-level telemetry.
1Configure & evaluate
Request API credentials or pull weights, then reproduce benchmark harnesses to validate MiniMax-M2 in your environment.
- Spin up vLLM, SGLang, or MLX with our starter configs.
- Run SWE-bench or BrowseComp notebooks to baseline behaviour.
- Document evaluation criteria before moving to staging.
2Instrument usage
Stream token, cost, and routing telemetry into your observability stack for live monitoring.
- Enable usage webhooks and budget alerts per workspace.
- Forward logs into your SIEM with <think> trace redaction controls.
- Track per-team cost trends to inform optimisation.
3Operationalise agents
Roll out tool-aware agents with guardrails, red-team them regularly, and refine prompts or adapters as workloads grow.
- Gate releases behind automated evaluation suites.
- Annotate interleaved thinking traces for policy review.
- Optimise decoding parameters per environment to balance cost and quality.
Why teams standardise on MiniMax-M2
Contrast closed APIs with an open, auditable MoE stack
MiniMax-M2 keeps quality high while giving you control over deployment, telemetry, and cost planning.
Black-box providers
Common blockers for agent teams
- Opaque reasoning traces make it hard to audit tool plans or debug production incidents.
- Token pricing changes without notice complicate runway planning.
- Limited region availability causes compliance and latency issues.
- Sparse insights into failure modes slow down QA and safety reviews.
- Restricted fine-tuning or retrieval rules limit domain adaptation.
- No direct access to weights prevents optimising for your hardware.
- Browser access and tool plugins vary across tiers, breaking consistency.
- Difficult to run red-team evaluations across coding, browsing, and finance scenarios.
MiniMax-M2 stack
Built for transparent, high-performance agents
- Open weights with interleaved thinking let you inspect and govern every response.
- Stable per-token pricing backed by real-time usage dashboards and webhooks.
- Global deploy options including self-hosting, managed platform, and local MLX builds.
- Benchmark-aligned evaluation harnesses for SWE-bench, Terminal-Bench, BrowseComp, GAIA, and more.
- LoRA and retrieval guides to tailor MiniMax-M2 to your domain while keeping safety signals intact.
- Telemetry surfaces expert routing stats and tool-call reasoning for compliance teams.
- Consistent tool schemas across shell, browser, retrieval, and Python agents.
- Community showcases like AnyCoder demonstrate real-world productivity gains with MiniMax-M2.
Launch-period free access plus predictable metered rates
MiniMax-M2 chat and API evaluations remain free for a limited window. Beyond the complimentary quota, usage bills at $0.50 per million input tokens and $1.50 per million output tokens with hourly metering and budget alerts.
Usage-based
MiniMax-M2 Pricing
Transparent pay-as-you-go billing for MiniMax-M2 with limited-time free access for hosted chat and API evaluation tiers.
$0.50per million tokens
$1.50per million tokens
Why teams choose MiniMax-M2 metered pricing
- Limited-time free access covers interactive chat and API evaluation while keep-alive workloads can upgrade seamlessly to metered billing.
- Input tokens bill at $0.50/M and output tokens at $1.50/M with hourly aggregation and budget alerts.
- Dashboards expose cost per workspace, persona, and project so finance stays informed.
Usage snapshots
- 750K input tokens consumed after the free tier → $0.38.
- 1M output tokens generated by a CI-driven repair workflow → $1.50.
- 3M mixed tokens for a multi-day browse-and-code investigation → roughly $2.40.
Hosted chat and API evaluation traffic remain free during the launch window; metered rates apply once usage exceeds the complimentary quota.
MiniMax-M2 FAQ
Answers to the questions teams ask before shipping
Ready to build with MiniMax-M2?
Download the MiniMax-M2 weights, open MiniMax Agent, or connect the Open Platform API to bring benchmark-grade agents into your stack.