MiniMax M3 (DeepInfra)

MiniMax released the M3 large language model on May 31, 2026, pairing frontier-tier coding and agentic performance with a 1-million-token context window and native multimodality, according to VentureBeat. The Chinese AI startup also said it planned to deliver M3 under an open-source license with open weights, allowing free enterprise downloads and customization. At launch, M3 was priced at $0.30 per million input tokens and $1.20 per million output tokens on fresh cache for the first week, with the full price set at $0.60 per million input and $2.40 per million output. VentureBeat reported that even at full price M3 costs a fraction of leading proprietary models, and the article claims it eclipsed GPT-5.5 and Gemini 3.1 Pro on selected benchmarks.

TypingMind published a setup guide for using the MiniMaxAI/MiniMax-M3 model ID through a DeepInfra API key, noting a June 1, 2026 release date, a 524,288-token context window and up to 512,000 max output tokens. The guide lists text, image and video input modalities with text-only output, and flags open weights alongside file upload, reasoning, tool calling and structured output support. Two configuration paths are documented: enabling the built-in DeepInfra provider and pasting an API key, or adding the model manually with the MiniMaxAI/MiniMax-M3 identifier and a 524,288 context length. The manual route sets the API type to the OpenAI Chat Completions interface, giving a concrete integration reference for teams wiring this variant into a third-party chat client.

MiniMax officially released MiniMax M3, a model that combines frontier-level coding and agentic performance with 1M-token context and native multimodality including image and video input plus computer operation. MiniMax claims M3 is the first and only open-weight model combining all three capabilities, which the company describes as table stakes for closed frontier models. M3 introduces MSA (MiniMax Sparse Attention), a new sparse attention architecture the company says partitions KV blocks more precisely than DSA or MoBA, avoiding the quadratic cost growth of full attention and making context a scalable dimension. MiniMax reports notable coding gains over M2 in bug fixing, frontend and backend development, and performance optimization, plus strong office-workflow agentic results. Access is offered through MiniMax Code, the Token Plan, and API services.

MiniMax positions M3 as an open-weight model combining three capabilities it says few models hold together: frontier coding and agentic work, million-token context, and native multimodality, with multimodal pretraining from step zero aligning textual and visual semantics. The vendor says the API supports up to 1M tokens with a guaranteed minimum of 512K, and the model page lists a separate M3.1 Flash entry. Vendor-reported results include a BrowseComp score of 83.5, which MiniMax claims exceeds Opus 4.7 at 79.3, plus a long-horizon demo in which M3 ran nearly 12 hours to independently reproduce an ICLR 2025 outstanding paper with 18 commits and 23 experimental figures. These figures come from MiniMax's own materials and were not independently verified in the supplied evidence.

MiniMax publishes open weights for MiniMax-M3, described as a native multimodal model with 428B total parameters and 23B activated parameters that supports a 1M-token context window. The official model card states that training was multimodal from the first step, giving deeper semantic fusion across text, image, and video rather than bolt-on visual handling. M3 relies on MiniMax Sparse Attention (MSA), a sparse attention operator tuned for million-token sequences that the card says cuts attention compute and memory footprint versus GQA while preserving quality. The card also claims 9x prefill and 15x decode speedups over M2 at 1M context and per-token compute reduced to 1/20, with a reasoning parameter offering enabled, adaptive, and disabled modes.

DeepInfra's API documentation for the MiniMaxAI/MiniMax-M3 model ID describes it as a native multimodal model with 1M context, 428B total parameters and 23B activated parameters. The page lists fp8 quantization, JSON and function calling, multimodal input support, a displayed 524,288-token context value, private endpoint deployment, and a Flex tier at 0.8x standard rate. The documented endpoint is OpenAI Chat Completions compatible, with support for simple messages, system prompts, multi-turn conversations, and streaming via SSE events terminated by a DONE marker. SDK and HTTP/cURL examples are provided, so clients already built on OpenAI-compatible interfaces can point at this model identifier without protocol changes.