Skip to main content
Back to timeline
arXivSource publication:

MiniMax-M2.7 activates about 9.8B parameters per token and scores 56.2 on SWE-bench Pro

Synopsis

MiniMax released the M2 series of Mixture-of-Experts language models, with the flagship M2 holding 229.9B total parameters while activating only 9.8B per token; the latest checkpoint, M2.7, scores 56.2 on SWE-bench Pro, in the same band as Claude Opus 4.6 at 57.3 and GPT 5.4 at 57.7. Most of the series' capability comes from agent-native post-training: data pipelines grounded in executable workspaces and artifact-aligned rewards, plus a reinforcement-learning system called Forge. M2.7 also shows an early form of self-evolution, autonomously triaging failed training runs and editing its own scaffold.

Interpretation

MiniMax built the M2 series as Mixture-of-Experts language models around the principle that mini activations can unleash maximum real-world intelligence, with the flagship M2 carrying 229.9B total parameters and activating only 9.8B per token. The prior route generally trades larger per-token compute for capability, whereas M2 places capacity in total parameters and holds compute down to the activated subset. Across agentic coding, deep search, office-task and reasoning benchmarks, the combination carried from M2 through M2.7 translates this small activation footprint into frontier-tier performance.

At each turn the model alternates reasoning segments with tool actions and keeps the previous turn's thinking blocks in context, so planning, acting and reflecting form one loop. The alternatives are front-loading all reasoning before any action, or stripping prior thinking each turn, which forces the model to re-derive context repeatedly. An ablation that removes thinking blocks from prior turns degrades agentic benchmarks consistently, with the largest gaps on multi-step tasks such as deep search and software engineering.

Post-training combines agent-driven data pipelines with Forge, a reinforcement-learning system that decouples training, inference and the agent itself and admits both white-box and API-only black-box agents. Long-horizon agent trajectories range from seconds to hours, so strict first-in-first-out scheduling is blocked by stragglers while fully greedy fetching pushes easy tasks into early batches. Windowed FIFO scheduling fetches out of order inside a window and enforces strict order across windows; in practice a window of 0.3 times the queue keeps near-FIFO distributional properties while substantially reducing cluster idle time.

Perspective

The report covers the M2 series' design, data pipelines, reinforcement-learning system and published benchmark scores; the benchmarks span agentic coding, agentic cowork and reasoning and knowledge, compared against Claude Opus 4.6, Claude Sonnet 4.6, GPT 5.4 and Gemini 3.1 Pro. The self-evolution account is limited to MiniMax's internal model iteration system and one 100-round autonomous scaffold optimization.

Self-evolution is described as an early form, supported by a 30% gain on in-house evaluations and by absorbing 30% to 50% of daily iteration workload. The hybrid sliding-window attention comparison shows clear deficits on long-context tasks, and the authors note that the correlation between proxy metrics and real downstream performance is fragile.