MiniMax M3

MiniMax (with authors from Peking University, NVIDIA, Zhejiang University, HUST, Nanjing University, Hangzhou Dianzi University) · 2026-06-11 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE efficient-attention multimodal open-weights frontier long-context

Design overview

MiniMax Sparse Attention (MSA) is the attention mechanism behind MiniMax M3, described here on the report's 109B/6B-active, 41-layer MoE testbed (3 dense + 38 MoE layers, 128 routed experts with top-4, 1 shared expert, 64 Q heads / 4 KV heads, head dim 128, 200K vocab, native text+image/video training). MSA keeps standard scaled dot-product softmax attention but makes its support learned and bounded: a lightweight Index Branch (one index query head per GQA group plus a single shared index key head) scores key-value blocks by max-pooled dot products and picks the top-k=16 of 128-token blocks for each query and group — the local block is always forced in — giving a fixed 2,048-token budget per query; the Main Branch then runs exact block-sparse attention over only those blocks. Because top-k selection is non-differentiable, the indexer is trained by a KL alignment loss against the group-averaged Main Branch distribution, confined to the index projections by a stop-gradient on its input, and stabilized by a two-stage warmup (full attention first, 40B tokens at this scale) plus the forced local block; ablations showed a learnable attention sink and an index value head are unnecessary, though a first-token sink emerges naturally. Efficiency comes from co-designed kernels: exp-free TopK in the small-k regime and KV-outer sparse attention with pre-scheduled chunking and a two-phase LSE combine for tensor-core utilization, delivering 14.2× prefill / 7.6× decode speedups on H800 at 1M context and a 28.4× attention-FLOPs cut. Two training routes are validated: from-scratch sparse pretraining (MSA-PT) and near-lossless conversion of a GQA full-attention checkpoint via 400B-token sparse CPT (MSA-CPT), plus a ~140B-token long-context extension evaluated at 128K. The production MiniMax-M3 model itself is released on HuggingFace but its configuration is not described in this report.

What the report claims is novel

Atlas assessment: MSA is the first learned blockwise-sparse attention record in the atlas: unlike linear-attention hybrids (kimi-k3's KDA, zyphra-zaya1, ai21-jamba) it keeps exact softmax attention and instead learns a per-GQA-group block support, and unlike the atlas' full-attention MoEs (deepseek-v3) it is sub-quadratic. Its closest published relatives — DeepSeek's DSA, NSA, MoBA, InfLLM-V2 — are not in the atlas, so MSA fills a genuinely new cell (learned sparse softmax attention on GQA). The per-group independent Top-k plus block-level selection and the KL+detach+warmup training recipe are the cleanest novel elements; the kernel work is engineering but substantial and open-sourced. Two caveats: all experiments are on the 109B model (the production 428B M3 config is not disclosed in this report, so scale-up behaviour is unverified here), and long-context quality still trails full attention slightly (HELMET-128K −0.60), with the paper itself noting a residual long-context retrieval gap.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyMiniMax M3[1]
OrganisationMiniMax (with authors from Peking University, NVIDIA, Zhejiang University, HUST, Nanjing University, Hangzhou Dianzi University)[2]
Release date2026-06-11
Report URLhttps://arxiv.org/abs/2606.13392[3]
Licensenot disclosed
Open weightsyes[4]

Variants

Scale

Total parameters109B[5]
Active parameters / token6B[6]
Layers41[7]
Hidden dimension3072[8]
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query heads64[9]
KV heads4[10]
Head dimension128[11]
Vocabulary size200000[12]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
10B pilot (ablation)10.53B1620481536324not disclosed

Core block

Block typesparse-MoE
Attention variantGQA[13]
Attention layer patternuniform: every layer uses MSA (all 41 layers; the 3 dense layers and 38 MoE layers share the same MSA attention module)[14]
Depth mixingsequential-residual
Activationnot disclosed
Expert count128[15]
Experts per token4[16]
Shared experts1[17]
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
MethodRoPE[18]
RoPE base frequencynot disclosed
Partial RoPEyes[19]
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksno[20]
Softcappingno

Context

Trained context lengthnot disclosed
Deployed context length1M[21]
Extension methodlong-context continued training: ~140B tokens of long-context training on top of the sparse-CPT checkpoint, evaluated on HELMET-128K and RULER-128K with the fixed 2,048-token attention budget preserved[22]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens3T[23]
Data compositionnative multimodal mixture of text and image/video data; detailed composition not disclosed[24]
Curriculum / stagingMSA-PT route: 40B-token indexer-warmup (full attention + KL) then sparse training for the remainder of the 3T budget; MSA-CPT route: starts from a GQA full-attention checkpoint trained on 2.6T tokens, replaces dense attention with MSA, and continues pretraining for 400B tokens (first 40B warmup, then sparse); plus a ~140B-token long-context extension stage[25]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTnot disclosed
Preference optimisationnone disclosed[26]
Reasoning trainingnone disclosed[27]
Distillationnot disclosed

Modality

Typemultimodal[28]
Attachmentnative[29]

Inference efficiency

KV-cache designGQA KV cache proportional to 4 KV heads; MSA sparsifies attention compute (top-16 of 128-token blocks = 2,048-token budget per query) rather than the cache itself — the report does not discuss cache-side savings
Quantisation shippednone disclosed[30]
Speculative decodingnone disclosed[31]
Serving optimisationsMSA co-designed GPU kernels (open-sourced at github.com/MiniMax-AI/MSA): exp-free TopK selection (per-thread register top-k with k-element min-heaps and shuffle merge; Bk=128, k=16), KV-outer sparse attention with query gathering and query concatenation for tensor-core utilization, pre-scheduled tile chunking for hot KV blocks, two-phase combine with per-partial LSE buffers (no atomics), LSE fusion into the forward pass, and dynamic load balancing via a persistent grid with atomic work claiming; measured 14.2× prefill and 7.6× decode wall-clock speedups on H800 at 1M context[32]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU (MSA-PT)67.2Table 2
MMLU-Pro (MSA-PT)38.8Table 2
BBH (MSA-PT)66.6Table 2
GPQA Hard (MSA-PT)26.3Table 2
GSM8K (MSA-PT)77.7Table 2
HumanEval (MSA-PT)64.0Table 2
EvalPlus (MSA-PT)61.8Table 2
BigCodeBench (MSA-PT)44.0Table 2
RULER-8K (MSA-PT)84.2Table 2
RULER-32K (MSA-PT)77.5Table 2
MMMU (MSA-PT)45.9Table 2
ChartQA (MSA-PT)75.4Table 2
VideoMME (MSA-PT)45.48Table 2
MLVU (MSA-PT)46.94Table 2
HELMET-128K Overall (MSA-CPT, after long-context extension)45.93Table 3
RULER-128K Overall (MSA-CPT, after long-context extension)72.12Table 3

Provenance references

  1. stated · Abstract ('A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3'); note: the report itself is titled 'MiniMax Sparse Attention' and its experiments use a 109B MSA model, not M3
  2. stated · Title page author affiliations
  3. stated · pipeline
  4. stated · Abstract ('A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3'); inference kernel open-sourced at github.com/MiniMax-AI/MSA
  5. stated · §5.1 ('approximately 109B total parameters')
  6. stated · §5.1 ('6B activated parameters per token')
  7. stated · §5.1 ('41-layer MoE backbone')
  8. stated · §5.1 ('hidden size d_model=3072')
  9. stated · §5.1 ('64 query heads')
  10. stated · §5.1 ('4 KV heads')
  11. stated · §5.1 ('head dimension 128')
  12. stated · §5.1 ('a 200K-token vocabulary'); also Appendix B.1 (pilot)
  13. stated · §2.3, §3.1 ('MSA, a GQA-based sparse attention mechanism'; 64 query heads, 4 KV heads, head dim 128)
  14. stated · §5.1 ('Each attention module uses MSA')
  15. stated · §5.1 ('128 routed experts')
  16. stated · §5.1 ('top-4 routed expert selection')
  17. stated · §5.1 ('1 shared expert')
  18. stated · §5.1 ('RoPE dimension 64')
  19. stated · §5.1 ('RoPE dimension 64' with head dimension 128 — RoPE applied to half of the head dims)
  20. stated · Appendix B.5 ('we do not include the learnable attention sink in the final recipe'); Appendix A (an attention sink on the first token emerges naturally in trained MSA models — emergent, not designed)
  21. stated · Abstract ('reducing per-token attention compute by 28.4× at 1M context'; '14.2× prefill and 7.6× decoding wall-clock speedups at 1M context length')
  22. stated · §5.3
  23. stated · §5.1 ('total budget of 3T tokens')
  24. stated · §5 ('a native multimodal model trained on a mixture of text and image/video data')
  25. stated · §5.1, §5.3
  26. stated · §7 (RL post-training is mentioned only as future work: 'extending the same selector-only design to settings beyond pretraining, including reinforcement-learning post-training')
  27. stated · §7 (mentioned only as future work)
  28. stated · §5 ('native multimodal model trained on a mixture of text and image/video data'; evaluation includes image (AI2D, ChartQA, MMMU, OCRBench v2, CharXiv, VisualWebBench, CVBench) and video (EgoSchema, LongVideoBench, MLVU, MMVU, VideoMME, TemporalBench) benchmarks)
  29. stated · Abstract, §5 ('native multimodal training', 'natively multimodal model'); encoder architecture not described in this report
  30. stated · §4.2 (kernels 'also support fp8' but no shipped quantization format is described)
  31. stated · report silent on speculative decoding
  32. stated · §4, Abstract, §5.4