MiniMax M2.5 / M2.7

Overview

MiniMax-M2.5 and MiniMax-M2.7 are the two most recent open-weights checkpoints of MiniMax's M2 series: sparse mixture-of-experts language models built, in the words of the series technical report, on the thesis that "mini activations can unleash maximum real-world intelligence," with a flagship configuration of 229.9B total parameters and only 9.8B activated per token (The MiniMax-M2 Series, arXiv:2605.26494). M2.5 was released on 2026-02-12 — the date MiniMax stamps on its launch announcement (MiniMax M2.5 announcement), matching the Hugging Face repository creation timestamp of 2026-02-12 (HF API metadata for MiniMax-M2.5) and the "MiniMax-M2.5 Version Release Date: 2026-02-13" in its model license (MiniMax-M2.5 LICENSE-MODEL). M2.7 followed on 2026-03-18 as a point release of the same series (MiniMax M2.7 announcement; MiniMax API release notes), shipped in two serving variants, MiniMax-M2.7 and M2.7-highspeed (MiniMax API release notes).

Positioning is squarely agentic and coding/office-work oriented. MiniMax describes M2.5 as "extensively trained with reinforcement learning in hundreds of thousands of complex real-world environments" and "SOTA in coding, agentic tool use and search, office work, and a range of other economically valuable tasks," headlined by 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench and 76.3% BrowseComp (MiniMax-M2.5 model card). Cost and speed are central to the pitch: M2.5-Lightning sustains 100 tokens/second at $0.30/M input and $2.40/M output, standard M2.5 runs at 50 tokens/second for half that, and MiniMax claims "$1 to run the model continuously for an hour" at 100 tok/s (MiniMax-M2.5 model card). M2.7 is pitched as "our first model deeply participating in its own evolution," able to build complex agent harnesses and complete elaborate productivity tasks using Agent Teams, complex Skills and dynamic tool search; MiniMax reports an internal M2.7 version autonomously optimizing a programming scaffold over 100+ rounds for a 30% performance improvement (MiniMax-M2.7 model card).

Architecturally the point release is a continuation, not a redesign: the series report treats the post-training pipeline as components that "co-evolve from M2 through M2.5 to the latest M2.7" on the same MoE base — a 62-layer decoder-only Transformer, 256 fine-grained experts with sigmoid gating, full multi-head attention with GQA, a 192K-token native context window and a Multi-Token Prediction module doubling as a speculative-decoding draft path (MiniMax-M2 series report HTML). Both released checkpoints report the same 229B model size on Hugging Face and identical config.json contents, including FP8 (float8_e4m3fn) block-wise quantization (M2.5 config.json; M2.7 config.json). Notable training-side techniques include the Forge agent-native RL framework (with a tree-structured sample-merging strategy MiniMax credits with ~40× training speedup) and the CISPO algorithm for MoE training stability (MiniMax-M2.5 model card), plus windowed-FIFO scheduling and prefix-tree merging in the report's system description (arXiv:2605.26494). Licensing differs between the two: M2.5's HF metadata declares license_name: modified-mit pointing at the GitHub LICENSE (HF API metadata for MiniMax-M2.5) with a "MINIMAX MODEL LICENSE" file in-repo (M2.5 LICENSE-MODEL), whereas M2.7 ships a "NON-COMMERCIAL LICENSE" that permits non-commercial use on MIT-style terms, requires "Built with MiniMax M2.7" attribution for commercial use, and prohibits any commercial use without prior written authorization from MiniMax (M2.7 LICENSE).

Architecture

Shared by both checkpoints unless noted. M2.5 and M2.7 publish byte-identical architecture fields in config.json.

• Total parameters: 229.9B for the series flagship per the technical report (arXiv:2605.26494); Hugging Face safetensors metadata reports 228,703,644,928 for M2.5 and 228,689,764,864 for M2.7, both surfaced as "229B params" on the cards (HF API metadata M2.5; HF API metadata M2.7; M2.7 model card).
• Activated parameters: 9.8B per token for the series flagship (arXiv:2605.26494); the report describes M2.7 as having "∼10B activated parameters" (MiniMax-M2 series report HTML).
• Architecture class: MiniMaxM2ForCausalLM, model_type: minimax_m2 — i.e. M2.7 reuses the M2 model class (M2.7 config.json; M2.5 config.json).
• Layers: 62 (num_hidden_layers: 62), a decoder-only Transformer (M2.5 config.json; MiniMax-M2 series report HTML).
• Hidden size: 3,072; expert intermediate size 1,536; no shared expert (shared_intermediate_size: 0) (M2.5 config.json).
• Attention: grouped-query attention with 48 query heads and 8 key-value heads, head dim 128, full (non-sliding) attention on every layer (attn_type_list is all 1s) (M2.5 config.json); the report confirms "full multi-head attention with GQA," 48 query / 8 KV heads (MiniMax-M2 series report HTML).
• QK normalization: use_qk_norm: true, qk_norm_type: per_layer (M2.5 config.json; M2.7 config.json).
• Positional encoding: RoPE with rope_theta: 5,000,000 and rotary_dim: 64 (partial rotary over the 128-dim head) (M2.5 config.json); the report lists Rotary Position Embeddings (MiniMax-M2 series report HTML).
• Experts: 256 routed experts (num_local_experts: 256), 8 activated per token (num_experts_per_tok: 8), sigmoid scoring with learnable routing bias (scoring_func: sigmoid, use_routing_bias: true) (M2.5 config.json); the report describes "256 fine-grained experts with sigmoid gating" and expert-specific bias terms (MiniMax-M2 series report HTML).
• Multi-token prediction: use_mtp: true, num_mtp_modules: 3, mtp_transformer_layers: 1 (M2.5 config.json). The report details a single MTP module (K=1) in pre-training expanded to three modules (K=3) during the continued-pretraining decay phase, MTP loss weight annealed 0.3 → 0.1, with the three modules used as speculative-decoding draft heads verified by the main model in one forward pass (MiniMax-M2 series report HTML).
• Precision / quantization: FP8 weights — quant_method: fp8, fmt: float8_e4m3fn, dynamic activation scaling, weight_block_size: [128, 128], with gate, e_score_correction_bias and lm_head excluded from conversion; base dtype bfloat16 (M2.7 config.json). HF lists tensor types F32 · BF16 · F8_E4M3 for both repos (M2.5 model card; M2.7 model card).
• Context window: max_position_embeddings: 196608 (192K) (M2.5 config.json); the report states a "192K-token native context window" reached via 8K → 32K → 192K context-extension stages (MiniMax-M2 series report HTML).
• Vocabulary: 200,064 tokens; tie_word_embeddings: false; activation silu; rms_norm_eps: 1e-06 (M2.5 config.json).
• Pre-training scale: 29.2T tokens total (19.9T constant phase, 9.3T decay-phase budget) (MiniMax-M2 series report HTML).
• Inference settings: temperature=1.0, top_p=0.95, top_k=40, with per-model default system prompts; served through SGLang, vLLM, Transformers (plus KTransformers for M2.5, NVIDIA NIM for M2.7) (M2.5 model card; M2.7 model card).
• Multimodality: text-in/text-out (pipeline_tag: text-generation for both repos) (HF API metadata M2.7); M2.7's MM Claw benchmark is described as an internal multi-modal office benchmark (MiniMax-M2 series report HTML).

Official vendor eval scores

Vendor-reported only. Unless noted, both columns come from Table 4 of MiniMax's own M2-series technical report, evaluated with thinking enabled and the interleaved-thinking protocol; AIME 2026, GPQA-Diamond, SciCode, IFBench, AA-LCR, HLE and MMLU-Pro follow the Artificial Analysis Index v4.0 protocol, no tools, pass@1 (MiniMax-M2 series report HTML).

Other vendor-reported figures: M2.5 scored 79.7 on SWE-Bench Verified under Droid scaffolding and 76.1 under OpenCode scaffolding, averaged 3.52M tokens and 22.8 minutes per SWE-Bench Verified task (vs 31.3 minutes for M2.1, a 37% speedup), and achieved a 59.0% average GDPval-MM win rate against mainstream models (MiniMax M2.5 announcement). M2.5's card appendix additionally reports AIME25 86.3, SciCode 44.4, IFBench 70.0 and AA-LCR 69.5 (M2.5 model card). For M2.7, MiniMax reports MM Claw 62.7%, 97% skill compliance across 40+ complex skills, MEWC v2 63.3, Finance Modeling Pro 57.0, RISE 64.3 and HyperTask 67.6 (M2.7 model card; MiniMax-M2 series report, Table 4).

Benchmark explanations

• SWE-Bench Verified — human-validated set of real GitHub issues; scores the share of issues whose generated patch passes the repository's tests (MiniMax evaluated it with Claude Code scaffolding averaged over 4 runs).
• SWE-bench Pro — industry-grade repository repair, a harder successor to SWE-bench aimed at realistic enterprise codebases (MiniMax-M2 series report).
• SWE-bench Multilingual — cross-language repository-level editing, testing coding competence beyond Python (MiniMax-M2 series report).
• Multi-SWE-bench — multi-repo task transfer: issue resolution spanning multiple repositories/languages (MiniMax-M2 series report).
• Terminal-Bench 2.0 — terminal and system-operation tasks in a sandbox (8 vCPU / 16 GB, 2-hour timeout, Terminus-2 scaffold in MiniMax's setup), measuring shell-agent reliability (MiniMax-M2 series report).
• NL2Repo — MiniMax's internal benchmark for synthesizing an entire repository from a natural-language specification (MiniMax-M2 series report).
• MLE Bench Lite — autonomous machine-learning engineering across 22 Kaggle-style competitions; scored as test-set medal rate after 24-hour runs on a single A30 (MiniMax-M2 series report).
• VIBE-Pro — MiniMax's internal end-to-end full-stack application development benchmark, verifying interaction logic and visual fidelity in containerized deployments (MiniMax-M2 series report).
• BrowseComp — hard open-web search benchmark measuring an agent's ability to locate obscure facts and synthesize an answer (MiniMax-M2 series report).
• Wide Search — broad-coverage web research benchmark requiring many parallel lookups aggregated into one answer, run in the WebExplorer agent framework (MiniMax-M2 series report).
• Toolathlon — heterogeneous tool-use benchmark exercising many different external tools in one agent workflow (MiniMax-M2 series report).
• GDPval-AA — the Artificial-Analysis-judged subset of OpenAI's GDPval suite of economically valuable office tasks, scored on deliverable quality (MiniMax-M2 series report).
• AIME 2026 — American Invitational Mathematics Examination problems; competition mathematics accuracy (MiniMax-M2 series report).
• GPQA-Diamond — hardest split of graduate-level physics/chemistry/biology questions written to be Google-proof (MiniMax-M2 series report).
• MMLU-Pro — broad multi-domain knowledge benchmark with harder, more distractor-rich multiple-choice items than MMLU (MiniMax-M2 series report).
• HLE (Humanity's Last Exam, no-tool subset) — frontier-difficulty expert questions across disciplines, answered without tools (MiniMax-M2 series report).
• tau2-bench — simulated-user tool-use benchmark with domain policies (retail/telecom splits) measuring policy-compliant multi-turn agent behavior; not scored for M2.5/M2.7 (n.a.) (MiniMax-M2 series report).
• LiveCodeBench — rolling competitive-programming benchmark built from recent contest problems to resist training contamination; not reported by MiniMax for this family (n.a.).
• RISE / MM Claw / MEWC v2 / Finance Modeling Pro / HyperTask — MiniMax-internal benchmarks for realistic multi-step browsing with a Playwright browser, multimodal office "claw" tasks, hard Microsoft Excel World Championship problems, expert-rubric-scored Excel financial modeling, and long-horizon multi-requirement "vibe coding" respectively (MiniMax-M2 series report).