Skip to content
AIBRAINWORX / THE INDEPENDENT AI LABRESEARCH · HARDWARE · OPEN INTELLIGENCE
Models & Weights / Inference & Hardware / Model Watch

MiniMaxAI Releases MiniMax-M2.5 Open-Weights Agentic Model for Coding, Search, and Office Workflows

MiniMaxAI published MiniMax-M2.5 on Hugging Face with reinforcement learning across 200,000+ environments, vLLM and SGLang serving guides, and modified-MIT licensing.

Listen to this articleBrowser Voice
Ready
AI-generated conceptual illustration for MiniMaxAI Releases MiniMax-M2.5 Open-Weights Agentic Model for Coding, Search, and Office Workflows; not a photograph or product screenshot
AI-generated conceptual illustration by AiBrainWorX; not a photograph or product screenshot.
Event / context: 2026-11-19. Source check: October 28, 2026.

MiniMax-M2.5 Targets Agentic Coding, Search, and Enterprise Workflows

MiniMax has published the weights and inference documentation for MiniMax-M2.5 on the MiniMaxAI/MiniMax-M2.5 model card on Hugging Face, positioning the model for autonomous software engineering, multi-step web research, tool calling, and structured office deliverables. Released on November 19, 2026 under the MiniMaxAI organization, the checkpoint reflects a broad shift toward reinforcement-learning pipelines designed for long-horizon execution rather than single-turn text generation. According to MiniMax, M2.5 was trained across hundreds of thousands of complex real-world environments to improve task decomposition and token efficiency across multi-step agent loops.

Alongside capability claims, MiniMax is pitching the release around low-cost continuous operation. The company states that running M2.5 continuously costs $1 per hour at a throughput of 100 tokens per second, or $0.30 per hour at 50 tokens per second. Developers browsing the repository today should also note a banner on the Hugging Face page indicating that a newer checkpoint, MiniMaxAI/MiniMax-M2.7, has since been released, placing M2.5 as a foundational milestone in the M2.x lineage rather than the newest entry in the family.

Reinforcement Learning Architecture, Benchmarks, and Serving Specs

Technical metadata on the Hugging Face repository identifies the model under the minimax_m2 architecture tag, distributed in Safetensors format with fp8 tagging and custom_code support. Loading the model directly through Hugging Face Transformers and accelerate requires enabling trust_remote_code=True so the custom modeling code can execute. In its release notes, MiniMax attributes the model's agentic behavior to its Forge Agent-Native RL Framework and dedicated agentic RL algorithm and reward design, trained across more than 200,000 real-world coding environments spanning over 10 programming languages, including Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, and Ruby.

In company-reported evaluations, MiniMax claims M2.5 scores 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp when paired with context management. MiniMax also reports completing SWE-Bench Verified 37% faster than its predecessor, MiniMax-M2.1, matching what the company characterizes as the speed of Claude Opus 4.6. To test generalization across different agent scaffolds, MiniMax evaluated M2.5 on SWE-Bench Verified using external harnesses, reporting 79.7 versus 78.9 for Opus 4.6 on Droid and 76.1 versus 75.9 on OpenCode. For full-stack application building, the team upgraded its VIBE evaluation to a Pro tier, where it claims overall parity with Opus 4.5.

For search and tool-calling tasks, MiniMax introduced RISE, short for Realistic Interactive Search Evaluation, alongside evaluations on BrowseComp and Wide Search. Across those three benchmarks, MiniMax reports that M2.5 achieves higher accuracy while using roughly 20% fewer search rounds than M2.1. On the deployment side, the model card documents OpenAI-compatible serving through vLLM and SGLang—including an lmsysorg/sglang:latest Docker workflow configured with --gpus all, --shm-size 32g, and --ipc=host—as well as Docker Model Runner, KTransformers, ModelScope, and links to community quantizations for llama.cpp, Ollama, and LM Studio.

Developer Workflow Impact and Domain Collaboration

For engineering teams, MiniMax highlights an emergent spec-writing habit in which M2.5 plans project structure, feature requirements, and user interface layout from the perspective of a software architect before emitting code. Rather than limiting output to isolated snippets or frontend demos, MiniMax states that the model supports 0-to-1 architecture setup, 1-to-10 core development, 10-to-90 feature iteration, and 90-to-100 testing and code review across Web, Android, iOS, and Windows applications, encompassing server-side APIs, business logic, and databases.

Beyond software development, MiniMax integrated M2.5 with its MiniMax Code interface, hosted API, and Model Context Protocol tooling to streamline multi-file and multi-tool workflows. To extend agentic reliability into knowledge work, MiniMax collaborated with senior practitioners in finance, law, and social sciences to define evaluation standards, provide feedback, and construct training data aimed at producing usable office deliverables rather than raw drafts.

Licensing, Security, Hardware, and Verification Caveats

Organizations considering M2.5 face several practical and governance caveats. First, the weights carry a modified-mit license rather than a standard unmodified MIT license, requiring legal teams to inspect the exact license terms before commercial redistribution or hosted integration. Second, the custom_code flag and trust_remote_code=True requirement mean engineering teams must audit the repository's Python implementation before loading weights inside restricted infrastructure.

Third, the public model card excerpt does not state total parameter counts, context-length ceilings, or exact VRAM requirements beyond fp8 tagging and multi-GPU Docker flags, leaving hardware planners to inspect the Safetensors index or quantized variants directly. Finally, all benchmark scores, speedups, and hourly operating cost estimates remain self-reported by MiniMaxAI—including publisher-built benchmarks like RISE and VIBE Pro—and training input from legal or financial specialists does not imply regulatory approval or guaranteed professional accuracy, particularly now that MiniMaxAI/MiniMax-M2.7 is listed as a newer alternative.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.

Follow your curiosity.

Pass the signal.