MiniMax Releases M3, an Open-Weight Model With a 1-Million-Token Context That It Says Tops GPT-5.5 on SWE-Bench Pro
Shanghai-based MiniMax launched M3 on June 1, pairing a 1-million-token context with a new sparse-attention design and company benchmarks that top GPT-5.5, with weights promised within 10 days.
Overview
Chinese AI start-up MiniMax debuted a new model, M3, on June 1, 2026, positioning it as an open-weight system aimed at long and complex coding tasks, according to the South China Morning Post. The model is billed as the first open-weight model to combine top-tier coding performance, a one-million-token context window, and native multimodality, The Decoder reported. On company-reported benchmarks, M3 edges out two leading proprietary models, though independent verification was not yet available at launch.
What We Know
M3 can process up to 1 million tokens of data at once, five times more than its predecessor, the M2.7, according to the South China Morning Post. The expanded context is delivered by a new architecture MiniMax calls MSA, or MiniMax Sparse Attention, which processes only relevant blocks of data rather than every pair of tokens, reducing compute costs, The Decoder noted.
On the coding benchmark SWE-Bench Pro, M3 scored 59.0%, surpassing OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro and approaching Anthropic’s Opus 4.7, as MarkTechPost reported. MiniMax also reported a 66.0% result on Terminal-Bench 2.1 and 70.06% on OSWorld-Verified, per MarkTechPost. On the BrowseComp test of autonomous web search, M3 scored 83.5, ahead of Opus 4.7’s 79.3, The Decoder reported.
The efficiency claims center on the sparse-attention design. At a 1-million-token context, MiniMax says M3 uses roughly one-twentieth of the per-token compute of the previous-generation M2, with prefill running more than nine times faster and decoding more than fifteen times faster than M2, according to MarkTechPost. The South China Morning Post separately reported that the architecture reduced computational requirements to as little as one-twentieth of previous levels, slashing inference costs while boosting response speeds, according to the South China Morning Post.
M3 is multimodal, accepting image and video input and capable of operating a desktop computer natively, and it was trained with mixed modalities from the start rather than having vision added afterward, The Decoder reported.
The API is live now, while the model weights and a technical report are scheduled to be released within 10 days of launch, with publication targeted for Hugging Face and GitHub, MarkTechPost reported. On pricing, requests up to 512,000 input tokens are billed at the standard rate while longer contexts cost more, and MiniMax’s subscription token plans range from $20 per month for roughly 1.7 billion tokens to $120 for 9.8 billion tokens, according to The Decoder.
What We Don’t Know
The benchmark figures are company-reported, and independent third-party scores were not yet available at launch. Because the weights and technical report had not been published when M3 debuted, the architecture and training details cannot yet be independently audited. MiniMax also did not disclose the size of the model or the computing infrastructure it used for training, according to the South China Morning Post.
Vendor-run benchmark claims warrant particular caution on SWE-Bench Pro specifically: The Machine Herald previously reported that the benchmark was found to contain systematic grading errors, underscoring why M3’s headline coding scores will need outside replication before they can be taken at face value.
Analysis
If the open weights arrive as promised, M3 would extend a year-long trend of Chinese labs shipping capable models under permissive terms, putting pressure on closed-weight incumbents not on raw benchmark leadership but on cost and deployability. The bet embodied in MSA is that attention-level efficiency, rather than brute-force scaling, is the path to making million-token context windows economical to run. Whether that bet holds will depend on the technical report and on independent benchmarks that, at launch, had not yet weighed in.