Model details
MiniMax-M3 represents the flagship entry in MiniMax's LLM lineup, combining frontier coding and agentic capabilities with native multimodality in a single model. Unveiled on June 1, 2026, it stands out as an open-weight release in a class where open alternatives have been rare. The model is marketed as achieving frontier-level performance on coding and agentic tasks while remaining fully accessible for self-hosting and commercial use, with distribution channels including Fireworks AI and Ollama's cloud offering. This positioning makes it attractive for teams that want strong development assistance without being locked into a closed-weight vendor.
At the architectural level, MiniMax-M3 introduces MSA (MiniMax Sparse Attention), a novel attention design that underpins its long-context behavior and differentiates it from the dense-transformer patterns of earlier MiniMax releases. The Fireworks launch coverage highlights MSA as the mechanism enabling extended context handling, while Ollama's library listing confirms practical integration with developer tools such as Claude Code, OpenCode, and Hermes Agent. The combination of native multimodality, sparse attention for efficient long-context inference, and open-weight availability positions MiniMax-M3 as a practical foundation for applications ranging from large-scale code understanding to multimodal agent workflows.
A provider subscription or plan supersedes token-based pricing for this model.
MiniMax Token Plan (minimax.io)
M3 reaches frontier capability on coding and agentic tasks, introduces the brand-new MSA (MiniMax Sparse Attention) supporting up to 1M context, and is a natively multimodal model. It is the only domestic model combining all three Frontier essentials and will be the only open-source one in this class.
MiniMax Token Plan (minimax.io)
Complete list of MiniMax latest updates for June 2026: get every product news, release note, and changelog from MiniMax summarized in one timeline.
MiniMax Token Plan (minimax.io)
Fireworks' Day-0 launch blog (June 12, 2026) provides detailed technical specifications for MiniMax-M3: open-weight release, 428B total / 23B active MoE architecture, native multimodality (text, image, video understanding), and a 500K-token context window at launch with the full 1M-token window promised in the coming d Fireworks launched M3 with full Day-0 support on its inference platform, framing the model as a frontier-capable open-weight option for long-horizon agentic workloads, visual-to-code pipelines, and production deployments requiring large context. The blog positions M3 alongside other 2026 open-weight releases (Kimi K2.5
MiniMax Token Plan (minimax.io)
Codersera's retrospective (last updated June 19, 2026) corroborates the MiniMax-M3 launch timeline: announcement on June 1, 2026, with weights published to the official MiniMaxAI Hugging Face repository by June 7, 2026. The article cites launch signals from researcher @arankomatsuzaki and MiniMax Head of Engineering @S The post aggregates concrete specs from the vendor model card: 428B total / 23B active MoE parameters, 1M-token context window, native multimodal input, API pricing of $0.30/$1.20 per million tokens, and a vendor-reported 59.0% score on SWE-Bench Pro (sourced via VentureBeat). For developers, it recommends M3 for open-
MiniMax Token Plan (minimax.io)
The official MiniMax platform release-notes page confirms that MiniMax-M3 was released on June 1, 2026 as the latest M-series language model, positioned for agentic reasoning, tool use, coding, multimodal chat input, and long-context tasks. This is the primary vendor source documenting M3's availability on the MiniMax While the excerpt does not disclose M3's architecture, parameter counts, or Token Plan–specific pricing, it establishes the model as the current flagship language model entry for agentic and coding workloads available via the MiniMax platform. Developers using the Token Plan can treat the June 1, 2026 release as the cu
This exact model name is also listed by 39 other providers.
