MiniMax's M3 just beat Opus 4.7 at browsing, trained itself, and never asked for help

Chinese AI startup MiniMax has released M3, a new flagship foundation model that the company claims is the first open-weight model to simultaneously deliver frontier coding performance, a million-token context window, and native multimodal understanding. The model is built on a proprietary MiniMax Sparse Attention (MSA) architecture and is available via API with automatic caching. It's a bet that open models can match closed rivals on hard, agentic tasks, not just chat benchmarks. See how M2.7 earlier surprised the field.

Benchmark highlights

MiniMax reports that M3 achieves industry-leading results across several coding and agentic benchmarks. On the BrowseComp agentic evaluation, M3 scored 83.5, surpassing OpenAI's Opus 4.7, which scored 79.3. The model also demonstrated strong performance on software engineering, terminal execution, and tool-use tasks. These results align with the broader trend that coding and browsing benchmarks are becoming the new battleground for frontier models, as noted in the race to own enterprise AI coding.

In an autonomous paper replication experiment, MiniMax tasked M3 with reproducing the ICLR 2025 Outstanding Paper 'Learning Dynamics of LLM Finetuning.' Over nearly 12 hours, M3 independently generated 18 commits and 23 experimental figures, successfully running the core experiments without human guidance. This raises a question the entire field is bumping into: how do you measure the real-world value of an agent's work? See the hardest problem in AI coding agents isn't code, it's counting the hours.

CUDA kernel optimization

M3 also demonstrated its autonomous engineering abilities by optimizing a FP8 matrix multiplication kernel on NVIDIA's Hopper architecture. Starting from only a task description and a non-functional Triton skeleton, M3 completed 147 benchmark submissions and 1,959 tool calls over roughly 24 hours, improving hardware utilization from 7.6% to 71.3%, a 9.4× speedup with zero human intervention. That's the kind of optimization that typically demands a team of kernel engineers weeks of work. It also surfaces a new challenge: as agents get better at optimizing code, verifying their output becomes harder. See why rewarding coding agents is getting harder.

PostTrainBench: M3 trains models

MiniMax also ran a test called PostTrainBench, where M3 was given four pre-trained base models and asked to autonomously complete the full post-training pipeline, data synthesis, training, evaluation, and iteration, within 12 hours. M3 scored 37.1, ranking third behind Opus 4.7 (42.4) and GPT-5.5 (39.3), but significantly ahead of all other tested models. That's a strong showing for an open-weight model, but the gap to the closed leaders is still real. For context on how agent training pipelines are evolving, see on-policy skill distillation boosts agent training without external memories.

Architecture and availability

M3 is built on the self-developed MiniMax Sparse Attention architecture, which supports an API context window of up to 1 million tokens, with a guaranteed usable length of at least 512K tokens. The model is natively multimodal, with text and visual semantic spaces aligned from the start of training, rather than through post-hoc patching. This kind of long-context, natively multimodal architecture is increasingly central to the open-weight ecosystem, as Gemma 4's encoder-free approach also demonstrates.

MiniMax positions M3 as the first model to bring a complete frontier capability set, coding, long-context agents, and native multimodality, into the open-weight ecosystem. The model is accessible via API, which includes automatic caching at no extra configuration. Whether M3 can sustain this lead as competitors like Meta and Google push their own open-weight families is an open question, but for now, MiniMax has drawn a clear line in the sand. See where this fits into the broader strategy in Alibaba's platform play against the model race.