MiniMax has finally released the full technical report for its M2 series, and the numbers are striking. The 35-page paper, posted on arXiv, details the architecture of the M2 through M2.7 models. The flagship M2 is a sparse mixture-of-experts (MoE) model with 229.9 billion total parameters, but only 9.8 billion are active for each token it processes. It uses 256 fine-grained experts, activating 8 per token.
The report covers the entire pipeline: architecture, data pipeline, the Forge training system, and the self-evolution mechanism of M2.7. MiniMax already released M2.7 back on March 18, 2026, with some benchmark results; this report fills in the technical details.
On the architecture side, M2 goes back to full attention (with grouped-query attention). The paper argues that linear or sparse attention perform okay on standard benchmarks but degrade on complex multi-step reasoning and tasks with contexts longer than 32k tokens. To reduce latency at the native 192k context length, the multi-token prediction (MTP) module predicts one future token during pretraining, then expands to three parallel tokens during fine-tuning via weight copying, serving as a draft path for speculative decoding. The routing layer ditches traditional softmax gating in favor of a sigmoid gating with learnable bias, removing the zero-sum competition between experts and smoothing activation.
The agent training platform, Forge, tackles throughput, stability, and flexibility bottlenecks. It uses a three-module decoupled design and introduces "windowed FIFO scheduling" to smooth computation blocking caused by sequence length variance. A "prefix tree merging" technique combines shared prefixes in multi-turn conversations, delivering up to 40x training speedups and significantly cutting memory consumption.
M2.7 demonstrates autonomous evolution: it can diagnose failed training runs on internal infrastructure, read logs, modify and upgrade its own scaffold, and iterate through multiple cycles. That self-improvement yields a 30% boost in internal evaluations. Benchmark results show M2.7, with its 9.8B active parameters, scoring 56.22% on the SWE-Pro software engineering agent benchmark, 52.7% on Multi-SWE-bench, and a 66.6% average medal rate across 22 competitions and three 24-hour autonomous runs on MLE Bench, matching Gemini-3.1.