Independent technical breakdown confirms MiniMax M3 is a 428B-parameter mixture-of-experts model with roughly 23B active parameters per token, paired with a 1M-token context window and native image and video input trained from step zero. The attention stack uses grouped-query attention combined with MiniMax Sparse Atte M3 scores 80.5% on SWE-bench Verified at official pricing of $0.30 per million input tokens, making it among the cheapest models above 80% on that benchmark and the only one in its class that reads image and video natively. It matches Claude Sonnet 4.6 on real-world agentic benchmarks. For UI automation and multimodal

Independent analyst commentary frames MiniMax M3's June 1, 2026 release as a landmark for open-weight models because it is the first to combine frontier-level coding performance, a 1M-token context window, and native multimodal input in a single architecture. The API went live the same day as the announcement, with ope The analysis flags an important credibility caveat: model weights were not available at launch, and MiniMax's benchmarks are vendor-reported rather than independently verified. While MiniMax is described as a Shanghai-based AI company founded in late 2021 by former SenseTime executives, the article argues that the real

The BenchLM benchmark aggregator tracks MiniMax M3 with a composite capability score of 63.9 out of 100, ranking it 49th of 232 models as of September 4, 2026. The model is classified as open-weight, non-reasoning, with a 1M-token context window. API pricing is $0.30 input and $1.20 output per million tokens, with cach Across 27 tracked benchmark rows, M3's strongest eligible category is Multimodal and Grounded, ranking 34th, with particular strength on screenshots, documents, charts, and grounded multimodal workflows. Category percentiles include Coding at the 63rd percentile and Knowledge at the 65th, while Agentic and Multimodal s

MiniMax officially released MiniMax M3 on June 1, 2026, as their flagship open-weight frontier model. The model uses a new attention architecture called MiniMax Sparse Attention (MSA), which the company describes as a clean, extensible sparse attention design that makes context another axis that can be scaled. It suppo On coding and agentic tasks, M3 shows significant improvements over the prior MiniMax M2, approaching leading closed-source models on bugfix, frontend/backend development, and performance optimization. It performs strongly on common office workflows such as search and Office-suite tasks and has become initially usable