Chinese Open-Weight Model 'MiniMax M3', Released with a Context Window of 1M Tokens

IBL News | New York

The Chinese model MiniMax M3, natively multimodal and with a context window of up to 1M tokens, was released this month. It’s an open-weight model, meaning its trained parameters are publicly available for download, local deployment, and fine-tuning.
However, the company withheld the original training code, training data pipelines, and specific inference operators. It can run on a desktop computer.

Users can use the model via MiniMax M3, MiniMax Code, the Token Plan, and API services.

According to the company,

“On SWE-Bench Pro, which measures coding capability, MiniMax M3 surpasses GPT-5.5 and Gemini 3.1 Pro and approaches Opus 4.7. On SVG-Bench, a benchmark that comprehensively evaluates SVG generation performance, MiniMax M3 surpasses Opus 4.7.”“On OmniDocBench, a multimodal benchmark, MiniMax M3 scores above Gemini 3.1 Pro. On Claw-Eval, an end-to-end evaluation framework for autonomous agents, MiniMax M3 achieves the highest score.”

“For long-horizon complex tasks, MiniMax Code’s Agent Team can break large tasks down into multi-stage, concurrent, and dynamically adjustable workflows, which are then advanced collaboratively by a cluster of agents. Through a Producer + Verifier adversarial harness loop, the Agent Team can continuously produce, reflect, and correct itself during execution. It can run autonomously for days without human intervention and ultimately deliver high-quality results.”

“We have seen that Claude Code has also recently released Dynamic Workflows in a similar direction. Compared with Claude Code’s stronger emphasis on fixed orchestration based on JS code, MiniMax Code focuses more on “deep reflection and continuous error correction”: the agent adjusts its plans and priorities in real time based on task progress, while users can step in at any time to add requirements or correct the direction.”

On the other hand, the company announced that MiniMax Code, “built on a harness based on the outstanding open-source community projects OpenCode and Pi”, will be open-sourced this project in the future.

IBL News | New York

Users can use the model via MiniMax M3, MiniMax Code, the Token Plan, and API services.

According to the company,

“On SWE-Bench Pro, which measures coding capability, MiniMax M3 surpasses GPT-5.5 and Gemini 3.1 Pro and approaches Opus 4.7. On SVG-Bench, a benchmark that comprehensively evaluates SVG generation performance, MiniMax M3 surpasses Opus 4.7.”“On OmniDocBench, a multimodal benchmark, MiniMax M3 scores above Gemini 3.1 Pro. On Claw-Eval, an end-to-end evaluation framework for autonomous agents, MiniMax M3 achieves the highest score.”

“For long-horizon complex tasks, MiniMax Code’s Agent Team can break large tasks down into multi-stage, concurrent, and dynamically adjustable workflows, which are then advanced collaboratively by a cluster of agents. Through a Producer + Verifier adversarial harness loop, the Agent Team can continuously produce, reflect, and correct itself during execution. It can run autonomously for days without human intervention and ultimately deliver high-quality results.”

“We have seen that Claude Code has also recently released Dynamic Workflows in a similar direction. Compared with Claude Code’s stronger emphasis on fixed orchestration based on JS code, MiniMax Code focuses more on “deep reflection and continuous error correction”: the agent adjusts its plans and priorities in real time based on task progress, while users can step in at any time to add requirements or correct the direction.”