CloudFerro Sherlock
Coverage
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…
$10 off the fastest Kimi K3 and GLM 5.2 from Synthetic
Model details
MiniMax-M2.5
MiniMax M2.5 is a text-only large language model in MiniMax's M-series lineup, sitting between earlier releases and the newer M2.7 successor that explicitly builds on its foundations. The model was published as open weights under the MIT license by the Shanghai-based MiniMax, making the full parameter set available for self-hosting and downstream research rather than restricting access to a hosted API. Across multiple distribution channels, M2.5 is positioned as a practical engine for software engineering and autonomous agent pipelines, with MiniMax's own product page framing it as state-of-the-art for coding and agent workloads designed for an "agent universe" of interconnected AI tools.
Training emphasis reportedly centers on reinforcement learning conducted in complex real-world environments spanning hundreds of thousands of machines, with the stated goal of producing efficient inference behavior and well-decomposed task plans that hold up under long agentic trajectories. In practice this means M2.5 is tuned less for casual chat and more for multi-step tool use, structured outputs, and code generation tasks that benefit from a large context window and stable reasoning across many turns. The model is distributed through several enterprise-grade channels, including an official NVIDIA NIM container on the NGC catalog under the minimax-ai organization and availability on Amazon Bedrock, so teams can choose between cloud-hosted inference and on-premises deployment depending on latency, cost, and data-residency needs.
CloudFerro SherlockMiniMaxAI/MiniMax-M2.5minimax
Quick Info
Powered by- Provider
- CloudFerro Sherlock
- Model key
- MiniMaxAI/MiniMax-M2.5
- Release date
- Mar 5, 2026
- Last updated
- Mar 5, 2026
- Knowledge cutoff
- 2026-01
- Input modalities
- Output modalities
- Capabilities
Cost
- Input token cost
- $0.30
- Output token cost
- $1.20
Limits
- Input tokens
- 180,000 tokens
- Output tokens
- 16,000 tokens
- Context window
- 196,000 tokens
Latest news about MiniMax-M2.5
CloudFerro Sherlock
CoverageRelease Notes
Discover more about what's new at AWS with Minimax M2.5 and GLM 5 models now available on Amazon Bedrock
CloudFerro Sherlock
Coverage
Chinese AI company MiniMax out of Shanghai has released its new open-weights model M2.5 under the MIT license.
CloudFerro Sherlock
CoverageBenchmark
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
CloudFerro Sherlock
Coverage
The Latent Space AINews roundup dated February 13, 2026 explicitly names MiniMax M2.5 and reports a claimed 80.2% score on SWE-Bench Verified, framed as matching Opus-level performance. This is the only candidate in the set that supplies a concrete technical data point for the M2.5 model, and the only piece of evidence The M2.5 mention is bundled into a broader news roundup led by Gemini 3 Deep Think, Anthropic's reported $30B raise at a $380B valuation, and GPT-5.3-Codex Spark, so M2.5 detail is limited to a single paragraph with no parameter count, context window, pricing, or tooling specifics. The 80.2% SWE-Bench figure is single-
CloudFerro Sherlock
CoverageRelease Notes
MiniMax's official Investor Relations news feed lists recent first-party releases spanning the M-series and H3 lines, including the M3 launch on May 31, 2026 (frontier coding, 1M-context MSA, native multimodality), MaxProof on June 9, 2026, the Shanghai International Film Festival partnership on June 28, 2026, the H3 o For developers tracking CloudFerro Sherlock's MiniMax-M2.5 listing, this IR feed is useful mainly as context: M2.5 is not the current MiniMax flagship, and newer M3/H3 posts dominate the news stream. The page provides no M2.5 release date, benchmark, context length, or architecture detail, so additional primary sources
CloudFerro Sherlock
CoverageBenchmark
SWE-Bench gap: 0.6%. Price gap: 10x. We ran both on the same real tasks. Here is the data and a one-screen decision matrix so your team can stop debating.