MiniMax-M2.7 (March 2026) is the latest release of the M2 series and keeps the architecture of the original M2 (October 2025): the M2, M2.1, M2.5 and M2.7 configs differ only in the position limit (196,608 → 204,800). It is a 229.9B-parameter MoE with 9.8B active, 62 blocks at width 3,072, and its defining choice is full causal attention in every layer: grouped-query attention with 48 query heads and 8 KV heads of dim 128, with no sliding window, no linear or sparse layers, and no attention sink or gate.
That was a deliberate reversal. MiniMax-Text-01 interleaved Lightning (linear) attention with softmax attention. For M2, MiniMax reports that such hybrids matched full attention on MMLU, BBH, MATH and LongBench but showed deficits in complex multi-hop reasoning at larger scale, that the proxy metrics it built for those deficits correlated only fragilely with downstream results, and that linear and sparse serving infrastructure still lagged: memory-bound kernels, precision-sensitive states, prefix caching and speculative decoding. Hybrid sliding-window variants, continued-pretrained at M2 scale for hundreds of billions to trillions of tokens, lost on long-context retrieval (RULER-128K common-word extraction 90.0 → 72.0) and in-context translation, and after SFT on long agent tasks (SWE-bench Verified 54.7 → 50.2), while matching or beating full attention on some tasks under 32K.
The price is KV memory: 2,048 cached values per token per layer, about 248 KiB per token in BF16 across the stack. The rest is conventional: 256 fine-grained experts with top-8 sigmoid routing and no shared expert, partial RoPE (64 of 128 dims, θ = 5,000,000), a QK RMSNorm taken over all heads at once rather than per head, and three MTP modules for self-speculative decoding, which are not included in the open checkpoints. MiniMax's next series, M3 (June 2026), adopted MiniMax Sparse Attention instead.