MiniMax-M2 series

MiniMax · 2026-05-26 · analysed 2026-08-10 · skill v0.4.0

sparse-MoE full-attention multimodal open-weights frontier frontier-generalist

Design overview

MiniMax-M2 is a 62-layer, 229.9B-parameter sparse MoE (9.8B active per token) built around the thesis that a mini activation footprint can reach frontier-tier agentic performance. Every block is a full-attention GQA layer (48 query / 8 KV heads, head dim 64, RoPE throughout) followed by an MoE FFN with 256 fine-grained experts, 8 active per token, routed by sigmoid gating with learnable per-expert bias terms that implicitly regulate utilization and let the auxiliary load-balancing loss be greatly reduced. The design is deliberately conservative on the mixing axis: after extensive hybrid sliding-window exploration (spanning ratios, RoPE variants, intra/inter-layer hybrids, and sink tokens), the team kept full attention in all 62 layers, arguing no sub-quadratic variant reliably matched full-attention quality in production settings — a counterpoint to the linear-attention hybrids in the same atlas generation. The 192K native context is reached by staged training (8K → 32K → 192K) rather than any positional-encoding extension. Capacity-per-token economics carry through to inference: a single MTP module (K=1, DeepSeek-V3-style, loss weight 0.3 annealed to 0.1) is expanded by weight copying to three draft modules (K=3) during the decay phase, giving multi-step speculative decoding, and the drafts are co-trained with the RL policy so acceptance survives distribution shift. The bulk of the report is agentic post-training: SFT on interleaved thinking traces, then CISPO agent RL over MDP-formulated trajectories with composite process/outcome/time rewards, run in the Forge system (windowed FIFO, prefix-tree merging up to 40× speedup, prefill-decode disaggregation), culminating in the M2.7 self-evolution checkpoint that debugs its own training runs.

What the report claims is novel

Atlas assessment: Block-level novelty is modest: fine-grained experts, sigmoid gating with learnable bias (aux-loss-reduced load balancing), and MTP-for-speculative-decoding all follow established DeepSeek-V3/DeepSeekMoE/aux-free lineages the report itself cites. Its distinctive position in the atlas is negative — a deliberate return to full attention at 230B scale while peers (moonshot-kimi-k3, upstage-solar-open-2, zai-glm-5-class hybrids) move to linear-attention hybrids; M2 keeps GQA + RoPE and buys 192K context with staged training rather than sub-quadratic layers. The genuinely novel contributions are training/infra: the Forge RL system (windowed FIFO, prefix-tree merging, white/black-box agent decoupling) and the M2.7 self-evolution loop, neither of which is a block mechanism. MTP weight-copy expansion (K=1→K=3) is a small but real refinement over DeepSeek-V3's single-module MTP.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyMiniMax-M2 series[1]
OrganisationMiniMax[2]
Release date2026-05-26
Report URLhttps://arxiv.org/abs/2605.26494[3]
Licensenot disclosed
Open weightsyes[4]

Variants

Scale

Total parameters229.9B[5]
Active parameters / token9.8B[6]
Layers62[7]
Hidden dimension3072[8]
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query heads48[9]
KV heads8[6]
Head dimension64
Vocabulary size200064[10]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
M2.5not disclosednot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed
M2.7not disclosednot disclosednot disclosednot disclosednot disclosednot disclosednot disclosed

Core block

Block typesparse-MoE
Attention variantGQA[11]
Attention layer patternuniform — full attention in all 62 layers (no hybrid, no sliding window in the final design)[12]
Depth mixingsequential-residual
Activationnot disclosed
Expert count256[13]
Experts per token8[14]
Shared experts0
Routingsigmoid gating with learnable expert-specific bias terms and top-8 selection (no softmax zero-sum constraint)[15]
Load balancinglearnable per-expert bias terms in the gating function, optimized jointly with model parameters, implicitly regulating expert utilization; auxiliary load-balancing loss greatly reduced (not eliminated)[16]
Expert granularityfine-grained: larger number of smaller experts with reduced per-expert FFN size (following DeepSeekMoE)[17]
MethodRoPE[20]
RoPE base frequencynot disclosed
Partial RoPEno
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksno
Softcappingno

Context

Trained context length192K[21]
Deployed context length192K[19]
Extension methodmulti-stage training curriculum: context window progressively extended from 8K through 32K to 192K during continued pre-training; no positional-encoding modification described[22]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens29.2T[23]
Data compositioncomprehensive curated corpus: web documents, academic literature, books, programming code, and structured question-answering content; quality assessed by model-based reward scoring plus auxiliary classifiers; balanced sampling upweights high-quality content while retaining category diversity; code, mathematics, and STEM significantly upsampled relative to their natural distribution[24]
Curriculum / stagingtwo-phase pre-training: constant phase on 19.9T tokens, then a decay phase with a 9.3T budget mixing short-text decay data with long-context data (high-quality code concatenation, naturally long-form PDFs, thematically related document packing); progressive context extension 8K → 32K → 192K; MTP loss weight annealed 0.3 → 0.1[25]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[26]
Preference optimisationother[27]
Reasoning trainingagentic RL over MDP-formulated trajectories (LLM as policy, context management/tool execution/memory as environment): CISPO policy gradient with composite rewards — process reward (language-mixing and tool-format penalties, structured-reasoning rewards), outcome reward, and time-efficiency shaping; interleaved chain-of-thought with reasoning-state persistence (full thinking blocks appended to message history across turns); multi-stage context-length curriculum in RL[28]
Distillationnot disclosed

Modality

Typemultimodal
Attachmentnot disclosed

Inference efficiency

KV-cache designGQA with 8 KV heads across all 62 full-attention layers: KV cache ∝ 8 heads and grows linearly with sequence length (no linear-attention layers to bound the cache)
Quantisation shippednone disclosed
Speculative decodingMTP-based speculative decoding: three MTP modules (K=3) generate draft tokens verified by the main model in a single forward pass; MTP modules co-trained with the RL policy via a top-K KL divergence loss to keep acceptance rates high under distribution shift; copy-initialized from main-model weights[29]
Serving optimisationsheterogeneous prefill-decode disaggregation with per-phase parallelism strategies optimized for MoE (eliminating prefill/decode mutual interference); inference kernels co-designed with the deployment stack inside the Forge RL system[30]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
SWE-bench Pro56.2Abstract, §7.5
SWE-bench Multilingual76.5Abstract, §7.5
Multi-SWE-bench52.7Abstract, §7.5
Terminal-Bench 2.057.0Abstract, §7.5
NL2Repo39.8§7.5
MLE Bench Lite (medal rate %)66.6§7.5
VIBE-Pro55.6§7.5
HyperTask67.6§7.5
MM Claw62.7Abstract, §7.3
BrowseComp77.8Abstract
GDPval-AA50.0Abstract
Toolathlon46.3Abstract
AIME 202694.2Abstract
GPQA-Diamond89.8Abstract

Provenance references

  1. stated · Title ('The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence')
  2. stated · Title page byline ([email protected])
  3. stated · pipeline
  4. stated · §7.4 ('The previous public M2 release (M2.5)'); §1 (within-series checkpoints released publicly; M2.7 is the current checkpoint)
  5. stated · §2.1 ('It contains 229.9B total parameters, with 9.8B activated per token')
  6. stated · §2.1
  7. stated · §2.1 ('a 62-layer decoder-only Transformer')
  8. stated · §2.1 ('a hidden dimension of 3,072')
  9. stated · §2.1 ('using 48 query heads and 8 key-value heads (GQA)')
  10. stated · §2.1 ('a vocabulary size of 200,064')
  11. stated · §2.1 ('full multi-head attention across all layers, using 48 query heads and 8 key-value heads (GQA)')
  12. stated · §2.1, §2.2.2 ('M2 adopts full multi-head attention across all layers'; hybrid SWA variants were explored and rejected)
  13. stated · §2.1 ('The MoE feed-forward layer contains 256 fine-grained experts')
  14. stated · §2.1 ('with 8 experts activated per token')
  15. stated · §2.1, §2.2.1 ('Routing is implemented using sigmoid gating with learnable expert-specific bias terms'); 'Instead of softmax-based top-k gating... we use sigmoid gating'
  16. stated · §2.2.1 ('Expert Bias... implicitly regulate expert utilization, allowing the auxiliary load-balancing loss to be greatly reduced')
  17. stated · §2.2.1 ('Fine-Grained Experts... a larger number of smaller experts, increasing the total expert count while reducing per-expert FFN size')
  18. n/a · §2.4
  19. stated · §2.1 ('a 192K-token native context window')
  20. stated · §2.1 ('Rotary Position Embeddings (RoPE) are applied throughout the model')
  21. stated · §2.1 ('pre-trained on 29.2T tokens with a maximum context length of 192K'); §2.4 (8K → 32K → 192K staged extension)
  22. stated · §2.4 ('Long-Context Extension... progressively extend the model's context window from 8K tokens through 32K and ultimately to 192K tokens')
  23. stated · §2.1 ('pre-trained on 29.2T tokens'); §2.4 (19.9T constant phase + 9.3T decay phase)
  24. stated · §2.4 ('Training Data', 'Data Distribution')
  25. stated · §2.4, §2.3
  26. stated · §5 ('We conduct Supervised Fine-Tuning (SFT) to instill the desired interleaved thinking behavior in M2'); SFT data interleaves thinking traces with intermediate actions and observations
  27. stated · §6.2.1 ('CISPO. We adapt Clipped Importance Sampling Policy Optimization (CISPO) [MiniMax-M1] to M2 series RL training'); clipped importance ratios with stop-gradient, zero lower bound, 1+ε upper bound
  28. stated · §6.2, §5, §7.1
  29. stated · §2.3, §6.3 ('MTP-based Speculative Decoding')
  30. stated · §6.3 ('Heterogeneous Prefill-Decode Disaggregation'), §1 (Forge)