TL;DR: MiniMax M3 is now available on Starchild. Its 1 million-token context window and agent-focused architecture make it a strong option for long tasks involving many steps, tool calls, files, and intermediate results.

Most models perform well when the task fits inside a short conversation. The harder test begins when an AI agent needs to inspect a repository, research across many sources, call tools repeatedly, and adjust its plan based on what comes back.

Those workflows generate a lot of context. If the model loses track of earlier instructions or discoveries, the agent starts repeating work, contradicting previous decisions, or drifting away from the original goal.

How long-running AI agents work

An AI agent typically operates as a loop:

1. Understand the objective and create a plan.
2. Call a tool, API, browser, or code environment.
3. Reason over the result.
4. Update the plan and take the next action.
5. Repeat until the work is complete.

Each cycle adds instructions, observations, tool outputs, code, and decisions to the working context. A larger context window allows more of that history to remain available as the agent moves through the task.

This matters for repository analysis, technical research, document review, data-heavy investigations, and other multi-step workflows where the early details can still affect decisions much later.

What MiniMax M3 changes

MiniMax M3 supports context windows of up to 1 million tokens using MiniMax Sparse Attention, an architecture designed to reduce the computing cost of processing very long sequences.

According to MiniMax, M3 uses roughly one twentieth of the per-token compute of its previous model at 1 million tokens. The company also reports more than 9 times faster prefill and 15 times faster decoding compared with MiniMax M2 at that context length.

The model was developed around coding and agentic work. In one MiniMax evaluation, M3 completed approximately 24 hours of continuous engineering work involving 1,959 tool calls and 147 benchmark submissions, using the feedback from each run to guide the next iteration.

That is the type of workload where a long-context model earns its place. Short prompts may feel similar across many capable models. The difference becomes more visible when an agent has accumulated hours of history and still needs to connect its latest result to a decision made near the beginning.

Built for the way Starchild agents work

Starchild agents handle work that grows over time. They can research across sources, inspect files, call APIs, write and run code, use connected tools, and continue through scheduled or multi-step workflows. Every action adds context that may affect what the agent should do next.

MiniMax M3 gives those agents more room to retain instructions, tool results, and earlier decisions as a task develops. This makes it a practical choice for repository analysis, ongoing research, document-heavy projects, and repeatable automations where continuity matters as much as the quality of any single response.

MiniMax M3 is available on Starchild now. Select it for an agent, give it a task with enough depth to test its planning and tool use, and see how well it holds the thread.

Try MiniMax M3 on Starchild