MiniMax M2.5
456B-parameter open MoE built for coding and agents — 80.2% on SWE-bench Verified at a fraction of frontier cost.
About
MiniMax M2.5 (February 2026) is the coding-and-agents flagship of the M series: a sparse mixture-of-experts with 456B total parameters, a 196K-token context window, and an Apache 2.0 license. It uses interleaved thinking — reasoning wrapped in <think> tags that must be preserved in conversation history — and was trained across 10+ programming languages and 200,000+ real-world environments, with Python as a primary language. It can emit 1,200+ lines of structured code in a single response (third-party report). It sits between M2.1 (Dec 2025) and M2.7 (Mar 2026, proprietary API-only) in a fast-moving line. Honest caveats from third-party testing: mandatory thinking adds ~2.2s latency on simple tasks, it can override system-prompt constraints when its thinking decides completeness matters more, and the free-tier output cap is a hard 8,192 tokens.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SWE-Bench Verified 80.2% — third-party research compilation (mzzkc, Apr 2026), not vendor-published
- Multi-SWE-Bench 51.3%, HumanEval 89.6%, LiveCodeBench 65.0%, Terminal-Bench 42.2% — same third-party compilation
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Release window (Feb 2026), 456B MoE, 196K context, Apache 2.0 verified across MagicHour, Medium deep-dive, and awesome-minimax; Benchmarks (SWE-bench 80.2% etc.) come from a third-party research compilation, not vendor-published — labeled as such in entry; Honest failure modes documented from third-party testing (thinking latency, system-prompt override, 8K output cap on free tier).