MiniMax M3: Open Multimodal Agents with 1M Context

MiniMax released M3 on June 1, 2026 as an open-weight model combining three capabilities that usually arrive separately: frontier coding and agents, a one-million-token context window, and native image/video understanding with computer use. The official release reports 59.0 on SWE-bench Pro, 66.0 on Terminal-Bench 2.1, 34.8 on SWE-fficiency, 28.8 on KernelBench Hard, and 74.2 on MCP Atlas.

The architecture contribution is MiniMax Sparse Attention, or MSA. MiniMax says it reduces per-token compute at one million tokens to one-twentieth of the previous model, with more than 9x faster prefill and 15x faster decode, while matching full attention across most ablations. Those are vendor measurements, but they address a real bottleneck: a nominal million-token window is not useful if processing it is prohibitively slow.

Release scope and open-weight status

M3 was available through MiniMax products and API on June 1, with the company promising weights and a technical report within ten days. The first-party model organization is the distribution point to confirm exact files and license. “Open-weight” is the defensible label; the complete pretraining corpus and every training recipe are not implied by the term.

The release also updated MiniMax Code, a scaffold co-trained for the model. Raw M3, M3 inside MiniMax Code, and M3 using Claude Code in a benchmark are distinct systems. The evaluation notes disclose that many coding scores used Claude Code with an overridden prompt, while other tests used specialized environments.

This disclosure is a strength. It prevents readers from assuming every result comes from one neutral harness. It also means a local deployment should not expect the score without rebuilding the agent layer.

Coding and agent benchmark table

The headline public results are:

The article contains many more internal and public evaluations. A compact table is safer than reproducing every chart because methodology varies by row. M3's Terminal value, for example, should not be compared to a “best reported harness” value without labeling the wrapper.

MSA makes long context computationally plausible

Full attention compares every query with every key and grows quadratically during prompt processing. Sparse attention adds a selection stage so each query reads a smaller set of blocks. MiniMax says MSA partitions the key-value history more precisely than alternative sparse approaches and uses a hardware-friendly gather order.

At one million tokens, the company reports per-token compute at 1/20 of M2.7, prefill over 9x faster, and decode over 15x faster. These figures describe an internal implementation and head configuration. Actual gains depend on sequence shape, accelerator, kernel, batch, and sparsity pattern.

Long context should be tested for both speed and retrieval quality. Place facts at different depths, add conflicting updates, and inspect whether the model follows the latest constraint. A large window can store a repository, paper, and logs, but it does not guarantee correct attention to them.

Native multimodality changes agent design

M3 was trained with interleaved modalities from the beginning. Image and video are not presented as a late adapter. The model can inspect figures, interfaces, and frames while maintaining the text and tool state of a long task. MiniMax uses a twelve-hour paper-reproduction case to illustrate this combination.

That case is impressive but anecdotal. The benchmark table offers more bounded evidence: Video-MME, OSWorld, OmniDocBench, MMMU-Pro, and related tests each expose a different slice. Frame rate and maximum frames are crucial. The release notes say external APIs sometimes accepted fewer frames, so raw scores do not always share identical input.

For computer use, isolate accounts and make sensitive actions confirmable. For video, test the sampling policy on rare events. For documents, require region citations and verify tables. One multimodal backbone simplifies context, but it also centralizes failure.

Long-running work is the intended product

MiniMax describes M3 as a collaborator for tasks that continue for hours. Its examples include reproducing a research paper and optimizing an FP8 matrix-multiplication kernel over 24 hours. In the kernel case, M3 reportedly made 147 submissions and 1,959 tool calls, raising hardware utilization from 7.6 to 71.3 percent.

Those demonstrations reveal the operational requirements: durable state, checkpointed work, bounded tools, cost caps, and objective verification. A twenty-four-hour agent without resumability or trace inspection is not production ready even if its final result is excellent.

The model should create intermediate commits or artifacts, report uncertainty, and stop at authority boundaries. Human reviewers need to inspect the trajectory without reading a million tokens. Summaries must link back to the underlying tool events.

Open deployment is a systems project

A one-million-token, native multimodal model imposes substantial memory and serving demands even with sparse attention. Teams need exact weights, quantization, engine compatibility, image/video preprocessing, and context limits. A checkpoint that loads is not yet a reliable service.

Measure prefill and decode separately across realistic prompt lengths. Track KV memory, batch interference, tail latency, and numerical quality after quantization. Pin tokenizer and chat templates. The open-model operations guide covers artifact provenance, evaluation gates, and rollback for this path.

API use avoids much infrastructure but changes the sovereignty and reproducibility story. Record which endpoint and service tier produced each trace. MiniMax's priority tier may change scheduling without changing model intelligence.

Where M3 is strongest and weakest

M3 is compelling for long repository work, multimodal research, computer-operating agents, and applications that need an open checkpoint rather than a hosted-only flagship. Its sparse-context engineering and broad benchmark program make it one of the strongest open-weight releases of the window.

It is less suitable for small devices, short low-latency classification, or teams unable to operate a large multimodal model. A million-token prompt can also be a security liability if untrusted files introduce hidden instructions. Ingest only authorized data and separate content from control instructions.

The model's internal VIBE, finance, and long-horizon tests should be treated as leads for local evaluation, not universal guarantees. Production evidence begins when the organization's own acceptance cases pass repeatedly.

Verdict and evaluation plan

Evaluate M3 at three levels: bare model, preferred open agent scaffold, and MiniMax's managed experience. Keep tasks, tools, and acceptance rules fixed. Measure resolution, long-context recall, visual grounding, tool errors, output tokens, wall time, accelerator cost, and reviewer effort.

Use SWE-bench-like repository cases, hour-scale interrupted tasks, documents and videos with known answers, and computer workflows that include protected actions. Require tests, citations, and human approval. The frontier model evaluation guide supplies the comparison manifest.

M3's release establishes that open-weight models can combine long context, agentic coding, and native multimodality. Whether it is the right production model depends on reproducing that combination under the chosen hardware and governance boundary.

Source notes — reviewed 2026

• MiniMax's M3 announcement supplies the date, architecture claims, evaluation values, and detailed methodology.
• The official MiniMaxAI model organization is the first-party weight distribution channel.
• MiniMax API documentation describes hosted access and service behavior.
• MiniMax's MaxProof report adds first-party evidence about M3's mathematical long-horizon evaluation.
• The official MCP Atlas repository documents one of the agent benchmarks cited by MiniMax.