MiniMax offloads part of M3.1 Flash inference to Inferact and strengthens its partnership with vLLM

MiniMax and Inferact are collaborating to leverage M3.1 Flash in production and integrate their optimizations into the vLLM open-source inference engine.

Publishing the weights of an artificial intelligence model is not enough to guarantee that it can be operated effectively by enterprises. Recent architectures impose constraints on memory, request management, and accelerator performance, to which inference engines must adapt. MiniMax and Inferact have established an engineering partnership to improve the deployment of their models in vLLM, while serving a portion of MiniMax M3.1 Flash traffic on Inferact's commercial infrastructure. Both companies are working on the compatibility of new architectures, decoding optimization, and performance validation as soon as the models become available. Improvements resulting from this collaboration are intended to be contributed back to the open-source project so that other developers can benefit from them. An inference infrastructure shared between two companies

MiniMax already uses Inferact's platform to serve a portion of the requests addressed to its M3 and M3.1 Flash models. Requests going through MiniMax's services can be routed to Inferact's infrastructure, alongside the compute capacities operated directly by MiniMax. The companies state that this production deployment was preceded by verifications regarding response quality, performance, and service reliability.

The operation relies on a separation between the model developer and the inference operator. MiniMax retains control over its model's service modalities, while Inferact provides compute capacities and an optimized environment to process requests. Its platform provides service level agreements, or SLAs, as well as automatic adjustments of execution parameters based on traffic evolution.

The value of this setup becomes apparent when demand fluctuates or when requests become more complex. An infrastructure must be able to accommodate new sessions, manage utilization spikes, and prevent certain operations from excessively slowing down others. In agentic applications, these constraints multiply because a single task can lead to numerous successive interactions with the model.

Both companies confirm the existence of this deployment, but do not disclose the share of traffic processed by Inferact, the number of requests involved, or the savings directly attributable to this setup. AI agents impose specific constraints on inference

M3.1 Flash specifically targets programming, process automation, and tasks involving multiple tools. A software development agent can analyze a code repository, inspect files, execute commands, examine an error, and then return to the model to select its next action. A single user request then turns into a sequence of mutually dependent requests.

These use cases place heavy demands on the inference engine. Contexts can lengthen over the course of interactions, certain portions of instructions are repeated, and generation alternates with calls to external tools. Wait times accumulate when each step must receive a response before the next one can begin.

Optimization is therefore not only about the number of tokens generated per second. The reuse of information retained in memory, request scheduling, long-input management, and costs associated with different accelerators also play a role.

MiniMax states that M3.1 Flash achieves generation throughputs exceeding 200 tokens per second under certain conditions, with a median exceeding 170 tokens per second, and claims a 70% reduction in usage costs compared to its predecessor. These figures pertain to the model's performance and pricing announced by MiniMax. They do not constitute a measurement of the gain achieved through the partnership with Inferact alone. Preparing vLLM ahead of new model releases

One of the partnership's goals is to ensure support for every new MiniMax release as soon as it is published. To achieve this, the teams work upstream on model-specific execution paths and verify their behavior on supported hardware configurations.

This work includes compatibility testing, decoding optimizations, and performance measurements. It becomes particularly important for Mixture-of-Experts architectures, in which not all parameters are activated for every token, as well as for multimodal models or those capable of processing long contexts.

A model published with accessible weights can be difficult to use immediately if the inference engine does not recognize certain operations, does not manage its memory properly, or requires specific adaptations. The work carried out before its release allows users to have a functional configuration available more quickly.

Contributions are intended for vLLM, an open-source inference engine from which Inferact originated and to which its teams continue to contribute. The collaboration thus benefits Inferact's commercial platform, as well as organizations operating vLLM on their own infrastructure. A commercial platform built around an open engine

Founded by creators and maintainers of vLLM, Inferact develops a commercial offering for enterprises seeking to operate models at scale. It combines the open-source engine with operational, monitoring, and optimization tools adapted to the constraints of a production service.

The company notably highlights the ability to adjust configurations based on available hardware and request profiles. A service handling long agentic conversations does not necessarily require the same settings as a classification API or a model used to rapidly generate short texts.

Inferact also works with accelerator vendors to improve engine performance. Its collaboration with AMD involves, for example, Instinct GPUs, ROCm libraries, and optimizations suited for long sessions. The company also collaborates with Google around running vLLM on TPUs.

For MiniMax, relying on such a platform allows it to outsource part of its operations while participating directly in the development of an open engine. For vLLM users, specific adaptations for MiniMax models can be reused independently of Inferact's commercial offering, subject to the