SiliconFlow says MiniMax M3 is now available through its OpenAI-compatible API, offering up to a 1 million-token context window, native multimodality, and launch pricing listed per million tokens. MiniMax describes M3 as an open-weight model aimed at coding, agentic tasks, long-context workloads, and multimodal use...
SiliconFlow said on June 2 that MiniMax M3 is now available through its OpenAI-compatible API, giving developers access to the model with a context window of up to 1 million tokens and launch pricing listed per million tokens.
The SiliconFlow post presents M3 as a model for coding, agentic workflows, long-context reasoning, and multimodal tasks. MiniMax separately announced M3 on June 1, describing it as a model that combines “frontier coding and agentic performance,” MiniMax Sparse Attention for long-context processing, and native multimodality.
According to SiliconFlow, MiniMax M3 can be used through an OpenAI-compatible API, which may make it easier for developers already using OpenAI-style request formats to test or integrate the model. SiliconFlow’s post lists launch pricing before discounts at $0.12 per 1 million cached input tokens, $0.60 per 1 million input tokens, and $2.40 per 1 million output tokens.
SiliconFlow also highlights benchmark claims for M3, including references to BrowseComp and SWE-Bench Pro. Those claims should be read as vendor-published performance statements unless independently reproduced, since the provided sources are from SiliconFlow and MiniMax rather than third-party evaluators.
MiniMax’s June 1 announcement describes M3 as a single model intended to cover coding, agentic behavior, long-context use, and multimodal understanding. The company says the model uses MiniMax Sparse Attention, or MSA, to support context lengths up to 1 million tokens.
On its product page, MiniMax describes M3 as the “first open-weight model” with coding and agentic frontier capability, 1M-context MSA, and native multimodality. The same product page says API support extends up to 1 million tokens, with a guaranteed 512K minimum context length.
The “open-weight” description is significant because it suggests MiniMax is positioning M3 differently from fully closed commercial models. However, the provided sources do not include the model’s license terms, weight distribution details, or usage restrictions, so the practical meaning of “open-weight” should be verified from MiniMax’s model release materials before deployment in production or redistribution.
A 1 million-token context window can be useful for workloads that require large volumes of source code, documentation, transcripts, research material, or multimodal inputs to be considered in a single request. MiniMax and SiliconFlow both emphasize long-context use as a major part of M3’s positioning.
Still, long context alone does not guarantee strong retrieval, reasoning, or task completion across an entire prompt. Performance can vary depending on the task, input structure, latency, and cost profile. SiliconFlow’s pricing makes the economics easier to estimate, but teams evaluating M3 will still need to test whether the model reliably uses information placed deep in long prompts.
MiniMax says M3 has native multimodality, while SiliconFlow’s post also describes the model as natively multimodal. The provided sources do not give enough detail to compare M3’s multimodal capabilities across image, text, audio, or video tasks, nor do they specify all supported input and output formats through SiliconFlow’s API.
For coding, SiliconFlow and MiniMax both frame M3 as a frontier coding model. SiliconFlow cites BrowseComp and SWE-Bench Pro benchmark claims, while MiniMax’s announcement emphasizes coding and agentic performance. Developers considering M3 for software engineering tasks should test it on their own repositories, build systems, and review processes, because benchmark results may not reflect local codebase complexity or production constraints.
SiliconFlow’s availability announcement gives developers another hosted route to MiniMax M3, with OpenAI-compatible access and clearly stated launch token pricing. MiniMax’s own materials position M3 as an open-weight, long-context, multimodal model for coding and agentic applications.
The key open questions are operational rather than promotional: how M3 performs in independent evaluations, how stable it is on very long prompts, what latency looks like at large context sizes, and what license terms apply to the open-weight release. For now, the sources establish that MiniMax has announced M3, SiliconFlow has made it available through its API, and both companies are positioning the model around 1 million-token context, coding, agentic tasks, and multimodality.
The SiliconFlow post presents M3 as a model for coding, agentic workflows, long context reasoning, and multimodal tasks.
MiniMax separately announced M3 on June 1, describing it as a model that combines “frontier coding and agentic performance,” MiniMax Sparse Attention for long context processing, and native multimodality.
SiliconFlow’s post lists launch pricing before discounts at $0.12 per 1 million cached input tokens, $0.60 per 1 million input tokens, and $2.40 per 1 million output tokens.
Continue reading
