MiniMax open-sources its weights so developers can use, improve, and optimize the model for wider distribution.
2
Serving a new model at launch requires custom kernels, benchmark work, and a continuing list of optimizations across attention, quantization, KV cache, and routing.
3
Long-horizon reinforcement learning tasks depend on carefully designed environments, data, rewards, and evaluations that distinguish real progress from hacking.
Summary
Olive Song describes MiniMax's reasons for releasing open-weight models and the post-training work behind MiniMax M3. The model is multimodal, supports applications such as computer use and game development, and was trained on text and image data from the start to avoid one modality collapsing. Song explains how reinforcement learning environments and reward design support difficult tasks such as kernel optimization and reproducing an ICLR paper over a twelve-hour run. Dan Fu explains what happens after a model is released. Together AI receives early architectural details, writes or adapts GPU kernels, and prepares an inference stack for day zero. Further work targets KV cache handling, attention, quantization, routing, and the needs of coding agents with long contexts and many tool calls. Fu expects GPU utilization and open models to improve substantially, while Song points to faster model development through self-evolution.
Open weights let developers improve the model and its serving stack
Olive Song says MiniMax releases its models because the open-source community is strong and can make direct contributions. Developers can use the weights, submit feedback and pull requests, and improve the model. Dan Fu gives a concrete example: Together AI can optimize the open-weight model so inference is faster and the service works better for users. Song connects this approach to MiniMax's mission of making intelligence available to everyone. The partnership also gives infrastructure teams access to new architectures and lets them prepare a service around the model before launch.
MiniMax M3 is used for multimodal agents, computer use, and game development
Song says MiniMax M3 differs from the M2 series because it is multimodal. It understands text, images, and videos, and builders are using it for multimodal agents. She mentions computer-use systems that navigate a computer and create things with tools. A less expected use is game development, where the model can help build games. The post-training recipe depends on the task. For kernel work, the team designs environments and data that let reinforcement learning optimize kernels and improve their performance through repeated attempts.
Fu describes Parallel Kernel Bench, a benchmark containing unsolved problems related to serving model inference. The team surveyed different inference approaches and found opportunities where no good kernels yet existed. He says benchmark overfitting is less worrying in this case because kernels that perform well on the benchmark can be taken and used to accelerate real inference. The benchmark therefore connects model-generated kernel work with practical serving improvements, rather than measuring performance on an isolated collection of already-solved operations.
A new model needs a working inference stack on launch day
When Together AI gets early details about a model such as MiniMax M3, the team studies architectural choices including sparse attention and mixture-of-experts decisions. It checks whether existing kernels work, modifies them when needed, or writes new kernels from scratch. Day zero is about delivering the expected quality and user experience. After launch, the team works through a list covering KV cache, attention kernels, quantization, and other components. Fu says the model should get faster from day zero to day seven and day fourteen as these optimizations land.
Agentic workloads change what the inference stack must optimize
Fu says inference design changes when workloads move from ordinary chat to coding agents that make hundreds of tool calls. These turn-based workloads affect KV cache behavior, prompting, routing, and kernel choices. A coding agent may upload an entire codebase, which creates a different context and optimization problem from a short system prompt followed by chat logs. The team watches how workloads evolve and adapts the inference engine to them. Fu also says lessons from optimizing one model, including sparse-attention work, can inform the next model even when the exact architecture differs.
Training multimodality from the start prevents one modality from collapsing
Song says M3 was trained multimodally from step zero, with both text and image data. She describes a problem seen in other labs where the model collapses after some multimodal training. MiniMax managed to avoid that failure. The team's attention maps showed text tokens attending to visual tokens, which allowed the modalities to combine naturally. This matters for tasks such as website development. A model can inspect how a website looks and then use that visual understanding to improve the site, including during reinforcement learning.
Long-horizon RL needs environments, rewards, and checks against hacking
Song describes training a model to reproduce an ICLR paper in a twelve-hour run. The task is difficult because it takes a long sequence of actions, requires GPUs, and has hardware constraints. MiniMax formulates the environment, data, rewards, and reinforcement learning algorithm so the task can be trained more efficiently. The model can submit several times, and the team evaluates each attempt. Because some attempts hack the evaluation, the team uses validation and testing to check whether performance really improved. Internal evaluations also come from using models to improve the team's own development work.
Long contexts make KV cache management resemble a distributed storage system
Fu compares handling KV cache for concurrent requests with very long contexts to rebuilding a distributed file system or a large database. The infrastructure must decide where the cache is stored, recognize when a prior computation exists, fetch it, and move it between locations. This becomes especially relevant for long-running agents and requests with contexts ranging from hundreds of thousands to around a million tokens. Fu calls the underlying ideas simple in theory, but says production systems still need to handle the storage and transfer work correctly.
"We do believe that the open source community as a whole is very strong and powerful."01:25
Who should watch
You are deploying an open-weight model and need to understand the work required before and after its public launch.
You are building coding or computer-use agents whose long contexts and repeated tool calls put pressure on inference infrastructure.
You are designing reinforcement learning for tasks that need many steps, external hardware, or evaluations that can be hacked.
