China's MiniMax just open-sourced a 1M-token model that outruns GPT-5.5 on real coding tasks

On April 24, 2025, MiniMax released M3, a model that breaks deliberately from its predecessor M2 and aims to redefine what an open-source frontier model can deliver. The timing matters. Western labs like OpenAI and Anthropic are walling off their best models behind paywalls and API restrictions. MiniMax chose to release M3 as an open-weight model, a decision that carries both technical and competitive weight.

The model bundles three things MiniMax calls necessary for a frontier system: native multimodality (image and video input), agentic coding, and a 1-million-token context window. All three live in a single architecture, not as post-hoc additions. This bundling, already done in models like GPT-4o and Claude 3.5, positions M3 as China's first open-source model to offer the full package.

The arithmetic of sparse attention

The technical foundation of M3 is MSA (MiniMax Sparse Attention), a new architecture designed to fix the core scaling limitation of full attention: quadratic complexity in sequence length. Sparse attention is not new, solutions like DSA and MoBA already exist, but MSA uses a more granular KV block partitioning method that the team says achieves higher effective context coverage.

The efficiency gains are real. At 1 million tokens, M3's per-token computation is 1/20th of M2's. The prefilling phase is more than 9x faster, and decoding enjoys more than 15x speedup. These numbers come from a custom kernel implementation, the KV outer gather Q, which reads memory sequentially and achieves a 4x speedup over open-source implementations like Flash-Sparse-Attention and flash-moba.

MiniMax reports that MSA's capability across most benchmarks matches that of full attention. This claim matters. If sparse attention introduces blind spots, losses in recall, hallucination, or instruction following, the context scaling gains come at a hidden cost. The benchmarks provided do not show regressions, but independent verification will tell.

A new approach to agentic training

MiniMax's treatment of agentic capability is the model's most consequential bet. Most coding benchmarks test single-turn tasks: a model receives a prompt and generates a solution. Real development is multi-turn, iterative, and collaborative. MiniMax built an interactive user simulator framework that trains the model on multi-turn interaction sequences, including requirement clarification, feedback revision, and task switching.

The results on standard agent benchmarks are solid: SWE-Bench Pro at 59.0%, Terminal Bench 2.1 at 66.0%, and MCP Atlas at 74.2%. But the framework itself matters more. By treating the agentic benchmark gap as a training data problem rather than an architecture limitation, MiniMax opened a path other labs can follow, one that does not require proprietary environments or closed infrastructure.

Real-world demonstrations: the difference between a benchmark and a workflow

Beyond benchmarks, MiniMax released two detailed case studies that show the model's capability in extended, autonomous execution.

Paper replication. MiniMax tasked M3 with independently reproducing an ICLR 2025 Outstanding Paper Award winner. The model ran for nearly 12 hours autonomously, producing 18 commits and 23 experimental figures. It matched the SFT phase probability trends, observed the squeezing effect discussed in the original DPO experiments, and validated the Extend mitigation method. All from a single input of the paper PDF and codebase, processed entirely within the 1M-token context window.

CUDA kernel optimization. The more technically revealing test involved an extremely narrow, high-specialty optimization problem: writing a production-grade FP8 GEMM kernel for NVIDIA Hopper GPUs. M3 received only a task description, a benchmark script, and a non-functional Triton skeleton. No reference code. Over 24 hours and 147 benchmark submissions, the model progressed through six rounds of optimization, from a baseline achieving 7.6% of hardware peak utilization to a final 71.3%. A 9.4x speedup. The optimal solution came at submission 145, after multiple plateaus. That persistence is rare among language models, which typically stop iterating after 30 attempts.

The PostTrainBench evaluation, where M3 trained four base models on math, tool calling, and coding tasks in 12 hours (scoring 0.37 against 0.42 for Opus 4.7 and 0.39 for GPT-5.5), suggests MiniMax sees agentic capability as a pipeline-level skill, not just a single-turn inference result.

The pricing strategy: undercutting by a factor of 15

The API pricing and Token Plan subscriptions make MiniMax's competitive positioning explicit. The Plus tier at ¥49/month ($6.80) delivers 600 million tokens, about five times the Claude Pro $20/month capacity. At extreme scale, the Ultra tier at ¥469/month ($65) offers 5.5 billion tokens, about three times the Claude Max $200/month capacity at roughly one-third the price.

MiniMax frames this as a rational move: make frontier models economically accessible to developers priced out by Western APIs. Whether this leads to a sustainable revenue model or a race to the bottom in Chinese AI cloud pricing depends on inference efficiency and demand scaling over the next year.

Complexities and open questions

M3's demonstrations are impressive, but several issues remain unresolved. First, the model's post-training data pipeline relies heavily on interleaved data, text, images, and videos mixed in training sequences. MiniMax's internal experiments show this scales better than synthetic data, but the team has not disclosed the exact proportion or curation methodology, making independent reproducibility difficult.

Second, the model is still being rolled out: the weights and technical report are promised within ten days. Until then, the community relies on API access and MiniMax's own benchmarks for verification, a trust requirement that open-source projects traditionally sidestep.

Finally, while M3's demonstrations show long-duration autonomy (12 and 24 hours), these are carefully scoped tasks. Whether the model can sustain reliable behavior across open-ended, ambiguous goals, the true test of agentic systems, remains to be seen.

A clear direction for open-source frontier models

MiniMax M3 does not claim to surpass GPT-4o or Claude Opus 4.7 across the board. But in bundling multimodal understanding, long-context reasoning, and agentic coding into a single open-source package, backed by a newly designed architecture and a training pipeline that rewards iterative interaction, it makes a strong case that open-weight models can compete on capability at a fraction of the cost.

The next step is adoption. If M3's open release and aggressive pricing attract a developer ecosystem that itself builds agentic workflows on top of the model, the gap between Chinese and Western frontier labs might narrow faster than many expect.