MiniMax-M2.5

BenchmarkList's leaderboard page explicitly names MiniMax-M2.5 and benchmarks it against peers including MiniMax Fable 5.1, Claude Fable 5.1, Claude Opus 5, Kimi K3, Qwen3.8 variants, and GLM 5.3 variants. The page reports an Agentic median percentile of 38th, with Tau2-Bench Telecom Success Rate at 95.3% (21/33, 94th The page is a third-party aggregator and includes cross-provider pricing columns (e.g., $0.12 in / $0.48 out for M2.5 alongside much higher figures for peers) that constitute serving-provider data rather than model news. Some benchmark entries contain LaTeX-style artifacts (\\\\method{}) suggesting scraping imperfectio

The release of MiniMax M2.7 adds enhancements to the popular MiniMax M2.5 model, built for agentic harnesses, and other complex use cases in fields such as…

Chinese AI company MiniMax out of Shanghai has released its new open-weights model M2.5 under the MIT license.

MiniMax, an AI company based in Shanghai, China, has announced the MiniMax M2.5, a frontier model designed to dramatically improve real-world productivity. M2.5 uses reinforcement learning in complex real-world environments of hundreds of thousands of machines to achieve efficient inference and optimized task decomposi

The Artificial Analysis provider page for MiniMax compares five MiniMax models—M3, M2.5, M2.7, M2.1, and M2—across intelligence, speed, price, latency, and context window metrics. Among open-weight MiniMax models, M2.5 scores a 23 Intelligence Index, ties with M2.7, and trails only M3 (29); it achieves 98 tokens/second The page contextualizes M2.5 within the broader MiniMax lineup, noting that for cost optimization M2.5 provides the most competitive pricing among the five tracked models, while M3 is highlighted as offering the best intelligence-plus-speed combination. The Artificial Analysis Intelligence Index v4.3.2 used for these s

The SemiAnalysis InferenceX page presents a technical architecture profile for MiniMax's M2 series covering both M2.5 and M2.7. It documents a 230B-total / 10B-active sparse MoE configuration built on the thesis that "mini activations can unleash maximum real-world intelligence," with 62 transformer layers using Groupe The page emphasizes an agentic, coding, and office-work orientation, citing MiniMax's positioning of M2.5 as "SOTA in coding, agentic tool use and search, office work, and a range of other economically valuable tasks," with headlines of 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp. M2.

Artificial Analysis provides a model page for MiniMax-M2.5, an open-weights MoE model with 230B total / 10B active parameters, a 205k token context window, and an MIT license. The page reports a $0.30 per 1M input and $1.20 per 1M output price (with a 90% cache discount noted as N/A), an output speed of 95.1 tokens per An Artificial Analysis Intelligence Index score of 23 is reported for M2.5, placing it above average among comparable open-weight models (median 18) in its class. Pricing is described as moderately priced on input but somewhat expensive on output relative to the medians of its peer group. The page also indicates a non-

BenchLM's catalog entry for MiniMax M2.5 assigns a composite capability score of 55.1/100, ranking it 90th out of 232 tracked models, with its strongest eligible category being Coding at rank 68 out of 151. The Agentic category ranks lower at 112 out of 152. Reported API pricing is $0.30 per million input tokens and $1 The page explicitly flags M2.5 as superseded by MiniMax M2.7, indicating it is no longer the current generation in the line. The reported context window of 128K tokens conflicts with creator-aligned sources citing 200k, suggesting either a different tracked configuration or stale aggregator data. As of September 10, 20

Discover more about what's new at AWS with Minimax M2.5 and GLM 5 models now available on Amazon Bedrock