Open-weight MoE model matching frontier coding results on SWE-Pro and Terminal Bench for long-horizon agentic workflows at low cost.
Modality:
Text
PDF
model ID
minimax/minimax-m2.7
Output Speed *
0.00
tok/s
Intelligence Index *
22.8
/ 100
Context Window *
196608
tokens
Input price
1.8
Anytoken
Output price
7.2
Anytoken
MiniMax M2.7: Frontier-Adjacent Coding and Agentic Execution at Open-Weight Economics
MiniMax M2.7 is a text-only, open-weight large language model from Shanghai-based MiniMax, built on a sparse Mixture-of-Experts architecture with roughly 230B total parameters and about 10B active per token. It is the flagship of the M2 series, positioned above M2.5 and M2.1 for complex coding and long-horizon agentic work. MiniMax M2.7 matches GPT-5.3-Codex on SWE-Pro and posts strong Terminal Bench 2 results, making it a fit for autonomous software engineering agents and multi-step productivity workflows rather than lightweight chat.
Start building with the MiniMax M2.7 API on AnyAPI.ai
Performance
Why M2.7 Wins on Real Software Engineering, Not Just Code Completion
MiniMax M2.7 targets production software engineering and long-horizon agents rather than single-turn generation. MiniMax reports 56.22% on SWE-Pro (matching GPT-5.3-Codex) and 57.0% on Terminal Bench 2, benchmarks that measure end-to-end task execution and system-level comprehension. Independently, Artificial Analysis places its Intelligence Index at 39, well above the ~29 median for comparable open-weight models. Because the MoE design activates only ~10B parameters per token, that capability comes at a low inference cost. For teams running high-volume coding agents or debugging pipelines, this combination keeps a capable model economically viable at scale.
Benchmarks
Independent Benchmarks: Strong Coding, Uneven Beyond It
Vals AI ranked M2.7 #10 on SWE-bench Verified — ahead of Gemini 3 Pro — and #12 on Terminal-Bench 2.0, beating predecessor M2.5. Artificial Analysis reports an Intelligence Index of 39 with concise output but slower-than-median generation, and a competitive time-to-first-token near 1.56s on MiniMax's API. However, Vals AI measured weak formal-math results (roughly 3% on Proof Bench) and mixed finance-agent scores. Interpretation: M2.7 is a coding-and-agent specialist. Treat its software-engineering strength as its defining property and validate carefully before math-heavy or finance-critical deployments.
Output Speed
*
0.00
tok/s
Intelligence Index
*
22.8
/ 100
MMLU *
Broad world knowledge and problem-solving
0
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
87
%
HLE *
Adherence to multi-step structured instructions.
30
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
0
%
Technical Specifications
What the model supports
MiniMax M2.7 is a text-in, text-out model with a 204,800-token context window that MiniMax defines as the combined total of input, reasoning, and output. Maximum completion is up to 196,608 tokens. Two most decision-relevant traits: reasoning (extended thinking) cannot be disabled on M2.x models, so every request incurs thinking overhead; and there is no image, audio, or video input. Function calling via tools/tool_choice and JSON-schema structured outputs are supported, making it suitable for tool-driven agent harnesses but not multimodal pipelines.
Verified Specifications —
MiniMax-M2.7
*
Input modalities
Text
PDF
output modalities
Text
Context window
196608
tokens
Maximum output tokens
176947
Reasoning
Yes
Knowledge cutoff
March 2026
Pricing (standard)
1.8
AnyTokens in
/
7.2
AnyTokens out
Comparison
MiniMax M2.7 vs MiniMax M3: Coding Specialist or 1M-Token Frontier?
Both models come from MiniMax's current lineup and share an agentic, tool-use orientation, so teams often weigh them directly. M2.7 is the open-weight coding-and-agent specialist with a 204,800-token combined context and text-only I/O. M3 is MiniMax's provider-designated frontier model, adding a 1,000,000-token context window and multimodal input. The practical decision is whether your workload needs frontier-scale context and multimodality, or whether M2.7's proven software-engineering performance and open weights better fit a cost-sensitive, self-hostable coding stack.
Dimension
MiniMax-M2.7
MiniMax-M3
Context window *
196608
tokens
524288
tokens
Output speed *
0.00
tok/s
91.67
tok/s
Intelligence Index *
22.8
29.2
Input pricing
1.8
AnyToken
1.8
AnyToken
Output pricing
7.2
AnyToken
7.2
AnyToken
Knowledge cutoff *
March 2026
June 2026
Choose MiniMax M2.7 when your priority is autonomous coding agents, debugging, and productivity workflows that fit within ~200K tokens, and when open weights, self-hosting, or low per-token cost matter. Choose MiniMax M3 when you need a 1M-token context for large codebases or long agent sessions, require multimodal input, or want the provider's designated frontier model. For most focused software-engineering agents, M2.7 delivers the stronger cost-to-capability ratio; M3 earns its place when scale and modality are hard requirements.
Limitations & Trade-offs
1
Always-on reasoning overhead. MiniMax documents that thinking cannot be disabled for M2.x models, so every request generates chain-of-thought before answering. This adds latency and token consumption even for simple calls, and Artificial Analysis measures M2.7 as slower-than-median for its size class. For low-latency, high-frequency, or trivial requests, a non-reasoning or lighter model is a better default.
2
Text-only, no multimodality. M2.7 accepts and produces text exclusively — no image, audio, or video input. Teams needing document-image parsing, UI screenshots, or diagram understanding must pair it with a vision model or choose MiniMax M3, which adds multimodal input. This is a hard limitation for any workflow depending on visual context.
3
Context smaller than frontier alternatives. The 204,800-token window is defined as the combined total of input, reasoning, and output, so a large completion reduces usable input. For whole-repository analysis or very long agent sessions, this falls short of 1M-token windows from MiniMax M3 and some competitors, and forces chunking or retrieval strategies.
4
Licensing and evaluation caveats. Reporting on M2.7's license is inconsistent — some sources cite MIT while MiniMax's own repository describes non-commercial terms requiring written authorization and a "Built with MiniMax M2.7" notice for commercial use. Verify the current license before commercial deployment. Independently, Vals AI also found weak formal-math and mixed finance-agent results, so validate outside core coding tasks.
Best-Fit Workloads
Where this model earns its place
01
Autonomous coding agents
M2.7's SWE-Pro score of 56.22% (matching GPT-5.3-Codex) and 55.6% on VIBE-Pro for end-to-end project delivery make it a strong engine for coding agents that plan, edit, and verify across Web, Android, and iOS tasks. Function calling and native Agent Teams support multi-agent harnesses. The always-on reasoning and slower generation mean it suits deliberate, high-value tasks more than latency-sensitive autocomplete.
02
Production debugging and incident response
MiniMax reports M2.7 correlating monitoring metrics with deployment timelines, performing causal reasoning on traces, and pinpointing root causes — in some internal cases reducing incident recovery to under three minutes. Its 57.0% Terminal Bench 2 result reflects system-level comprehension useful for log analysis and SRE-style workflows. Pair with real tool access to databases and repositories for the strongest results.
03
Office document generation and editing
M2.7 produces and performs high-fidelity multi-round edits on Word, Excel, and PowerPoint deliverables, and holds the highest open-source GDPval-AA ELO (1495). This suits automated report generation, financial modeling drafts, and structured content pipelines. Because output is text-only, downstream rendering into final office files is handled by your application layer rather than the model directly.
04
High-volume agentic workflows on a budget
With only ~10B active parameters per token, inference cost and latency track the active count, giving M2.7 a favorable cost-to-capability ratio for teams running many agent invocations. It occupies a low-cost tier relative to frontier proprietary coding models while retaining competitive coding accuracy — well suited to cost-sensitive, high-throughput automation where a frontier model would be prohibitively expensive at scale.
Pricing in anytokens via AnyAPI
Input
1.8
₳
Output
7.2
₳
Cache write
—
₳
Cache read
0.36
₳
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04