Wait Which Model?
← Back to directory
MiniMax

MiniMax-M2.5

Released Feb 12, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Text
Context window
205K
Max output
—
Speed
95 tok/s · 1.67s to first answer token
Price ($/MTok in / out)
$0.15 / $1.20
Cost per task
—
Open weights
Yes
Benchmarks4 of 8 reported

GPQA Diamond

85.2%

SWE-bench Verified

80.2%

AIME (latest)

86.3%

Humanity's Last Exam

19.4%

MMLU-Pro

—

Terminal-Bench 2.1

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • SWE-bench Verified jumped sharply over M2 in under four months, the fastest capability gain in the series so far
  • Cheap enough to run continuously — MiniMax markets a roughly $1/hour cost at 100 output tokens/sec
  • A separate 'Lightning' throughput mode trades some cost efficiency for materially faster output when latency matters more than price

Weaknesses

  • Still text-only and capped at the same ~205K context as M2 — the long-context and multimodal jump waited for M3
  • MMLU-Pro wasn't part of MiniMax's own benchmark table at launch, an unusual omission next to its detailed coding and agentic suite
  • Superseded within four months by M3, the usual pace for this series

Officially announced 2026-02-12 (minimax.io/news/minimax-m25); MiniMax's own release material frames this as an iteration on M2/M2.1 rather than an explicit replacement of either, so predecessorId is left null. GPQA-Diamond 85.2, AIME25 86.3 and HLE 19.4% (without tools) are from the official Hugging Face README's comparison table; SWE-bench Verified 80.2%, which MiniMax describes as a new record for the series, is from the same announcement. MMLU-Pro is absent from MiniMax's own table; a third-party 74% figure (oddly lower than M2's official 82) could not be corroborated and looks inconsistent with M2.5's gains elsewhere, so left null. Context window confirmed at 204,800 tokens via MiniMax's own API docs — the same as M2; the 1M-token window is M3-only, despite some third-party trackers conflating the two. Pricing shown is the standard (non-Lightning) tier; the Lightning variant runs faster (100 tok/s vs 50) at roughly double the output price ($0.30/$2.40 vs $0.15/$1.20).

For developers

API model strings

  • minimaxMiniMax-M2.5

Licence

MiniMax Modified MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News