MiMo-V2

Xiaomi (LLM-Core) · 2026-01-06 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE efficient-attention text-only open-weights frontier reasoning-specialised

Design overview

MiMo-V2-Flash is Xiaomi's 309B-parameter sparse MoE (15B active per token) built around a deliberately aggressive hybrid attention schedule. Eight hybrid blocks each stack five sliding-window-attention layers (128-token window, 64 query / 8 KV heads) with one global-attention layer (64 query / 4 KV heads), giving 39 SWA and 9 GA layers; only the very first block is global with a dense FFN. A learnable attention sink bias in the softmax denominator — following OpenAI's gpt-oss — lets each head dump surplus probability mass, which the authors show is what lets the 5:1 ratio and 128-token window match or beat an all-global baseline on general, long-context, and reasoning benchmarks. Position comes from partial RoPE (first 64 QK dimensions); the GA RoPE base is raised from 640,000 to 5,000,000 when the context is extended from 32K to 256K, while SWA keeps base 10,000. Every MoE layer routes each token through 8 of 256 experts (no shared experts, per-expert width 2048) under a sequence auxiliary loss plus expert-bias updates. Training runs FP8 mixed precision over 27T tokens in three stages (general at 32K, code/reasoning-heavy mid-training, 256K context extension). The headline post-training contribution is MOPD, multi-teacher on-policy distillation: domain-specialized RL teachers supply dense token-level reverse-KL rewards combined with GRPO-style outcome rewards so one student absorbs every teacher's peak capability. The lightweight 0.33B MTP blocks (dense FFN + SWA) double as a self-speculative draft model — ~3.6-token acceptance length, 2.6x decoding speedup — and are open-sourced alongside the weights.

What the report claims is novel

Atlas assessment: The block is a competent synthesis of known mechanisms rather than a new mixing family: the 5:1 SWA/GA schedule extends the Gemma-style local/global interleave, the learnable attention sink bias is taken from OpenAI gpt-oss (in atlas), and FP8 mixed precision + MTP + GRPO-style ORMs follow the DeepSeek-V3/R1 playbook. The genuinely distinctive contributions are (a) the aggressive 128-token window at 5:1 validated with sink bias, and (b) MOPD, a named multi-teacher on-policy distillation paradigm — the same acronym and concept later appear in Moonshot Kimi K3 (arXiv 2607.24653, in atlas), which postdates this report; MiMo's January 2026 publication predates it, so the atlas now holds two independent MOPD claims. The 'rivals DeepSeek-V3.2/Kimi-K2 at 1/2-1/3 parameters' claim is internal and not independently verifiable. Report is unusually silent on normalization, activation function, router function, vocab size, and training hardware — architecture sections that peers almost always fill.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyMiMo-V2[1]
OrganisationXiaomi (LLM-Core)[2]
Release date2026-01-06
Report URLhttps://arxiv.org/abs/2601.02780[3]
Licensenot disclosed
Open weightsyes[4]

Variants

Scale

Total parameters309B[5]
Active parameters / token15B[6]
Layers48[7]
Hidden dimension4096[8]
FFN inner dimension2048[9]
FFN ratio0.5
Query heads64[10]
KV heads8[11]
Head dimension192[12]
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention varianthybrid[13]
Attention layer pattern8 hybrid blocks of 5 SWA + 1 GA; the very first Transformer block is GA with a dense FFN; totals 39 SWA + 9 GA layers; SWA window = 128 tokens[14]
Depth mixingsequential-residual
Activationnot disclosed
Expert count256[15]
Experts per token8[16]
Shared experts0[17]
Routingnot disclosed
Load balancingsequence-wise MoE auxiliary loss (coefficient 1.0e-5 all stages) plus per-expert bias updates (bias update factor 0.001 in Stages 1-2, 1.0e-5 in Stage 3; 1.0e-4 during SFT)[18]
Expert granularitynot disclosed
MethodRoPE[21]
RoPE base frequency5000000[22]
Partial RoPEyes[23]
Typenot disclosed
Placementnot disclosed
QK-normno
Attention sinksyes[24]
Softcappingno

Context

Trained context length262,144[25]
Deployed context length262,144[26]
Extension methodStage-3 context-extension training (26-27T) at 262,144 with long-range data upsampled; GA RoPE base raised 640,000 -> 5,000,000[27]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens27T[28]
Data compositiondiverse high-quality corpus: public web content, books, academic papers, code, mathematics, and broader STEM; deliberate shift toward long-range dependencies (long-form web documents, repository-level code, pull requests, issues, commit histories); ~5% synthetic reasoning data added in Stage 2; pipeline largely follows MiMo-7B[29]
Curriculum / stagingthree stages: Stage 1 (0-22T) general corpus at 32K context; Stage 2 (22-26T) code upsampling + ~5% synthetic reasoning data; Stage 3 (26-27T) context extension to 256K with long-range data upsampled[30]
OptimizerAdamW with beta1=0.9, beta2=0.95, weight decay 0.1, gradient clipping max norm 1.0[31]
LR scheduleStage 1: linear warmup 0 -> 3.2e-4 over 50B tokens, constant 3.2e-4 for 12T, cosine decay to 1.0e-4 over 10T; Stage 2: cosine 1.0e-4 -> 3.0e-5 over 4T; Stage 3: cosine 3.0e-5 -> 1.0e-5[31]
Batch schedulebatch size warms up linearly to 2048 over the first 500B tokens and stays constant through Stages 1-2; fixed batch size 256 in Stage 3; SFT batch size 128[18]
PrecisionFP8 mixed precision (DeepSeek-V3-style): BF16 for attention output projections and embedding/output head, FP32 for MoE router; FP8 used for both training and inference[32]
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[33]
Preference optimisationGRPO[34]
Reasoning traininglarge-scale RL: non-agentic RL on verifiable domains (programmatic verifiers + LLM judge, rubric-based judge for subjective qualities); agentic RL across code (120K environments), terminal (~30K), web development, search (150K), function-calling (50K) agents; all consolidated via MOPD distillation[35]
DistillationMulti-Teacher On-Policy Distillation (MOPD): student samples from its own distribution and receives dense token-level reverse-KL rewards from domain-specialized teacher policies (RL or SFT trained), combined with ORM advantages; training-inference importance sampling with token discarding (IcePop2025)[36]

Modality

Typetext-only
Attachmentn/a[37]

Inference efficiency

KV-cache designhybrid: 39/48 layers are SWA with a 128-token window so their KV cache is bounded by the window; 9 GA layers keep full cache; report claims nearly 6x reduction in KV-cache storage and attention computation for long contexts[38]
Quantisation shippedFP8 (used for both training and inference)[39]
Speculative decodingself-speculative decoding with MTP as native draft model: single MTP head during pre-training, replicated to K=3 jointly-trained heads in post-training; up to 3.6 average acceptance length and 2.6x decoding speedup[40]
Serving optimisationsnone disclosed[41]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU-Pro84.9Table 9
GPQA-Diamond84.3Table 9
HLE (no tools)22.1Table 9
AIME 202594.1Table 9
HMMT Feb. 202584.4Table 9
LiveCodeBench-v685.1Table 9
Arena-Hard (Hard Prompt)54.1Table 9
Arena-Hard (Creative Writing)86.2Table 9
LongBench V260.6Table 9
MRCR45.7Table 9
SWE-Bench Verified73.4Table 9
SWE-Bench Multilingual71.7Table 9
Terminal-Bench Hard30.5Table 9
Terminal Bench 2.038.5Table 9
BrowseComp45.4Table 9
BrowseComp (w/ Context Manage)58.3Table 9
tau2-Bench80.3Table 9
NIAH-Multi (256K)96.7Table 6
GSM-Infinite Hard (128K)29.0Table 6
MMLU (Base)86.7Table 5

Provenance references

  1. stated · Title (MiMo-V2-Flash Technical Report); Table 9 header splits 'MiMo-V2' / 'Flash'
  2. stated · Byline 'LLM-Core Xiaomi'
  3. stated · pipeline
  4. stated · Abstract: 'We open-source both the model weights and the three-layer MTP weights'; §1 (github.com/XiaomiMiMo/MiMo-V2-Flash)
  5. stated · Abstract, §3.2, Table 5
  6. stated · Abstract, §3.2, Table 5 (#Activated Params 15B)
  7. stated · Table 1 (Layers Total/SWA/GA = 48/39/9), §3.2
  8. stated · §3.2 ('The hidden dimension is set to 4096')
  9. stated · §3.2 ('an intermediate hidden dimension of 2048 for each expert'); dense first-layer FFN intermediate is 16384
  10. stated · Table 1 (SWA Heads (Q/KV) 64/8; GA Heads (Q/KV) 64/4)
  11. stated · Table 1
  12. stated · Table 1 (Head Dimensions (QK/V) 192/128 for both SWA and GA)
  13. stated · §2.1 (interleaved Sliding Window Attention and Global Attention, both GQA)
  14. stated · §2.1, Table 1
  15. stated · §2.1 ('Each MoE layer comprises 256 experts in total'), Table 1
  16. stated · §2.1 ('8 activated per token'), Table 1 (Experts (Total/Activated) 256/8)
  17. stated · §2.1 ('and contains no shared experts')
  18. stated · §3.2, §4.2
  19. stated · §3.2, §3.1 Stage 3
  20. stated · §3.1 Stage 3 ('extend the model's context window to 256K tokens')
  21. stated · §2.1 ('Rotary Positional Embedding (RoPE) is partially applied to the first 64 dimensions of query and key')
  22. stated · §3.2 (Long Context Extension)
  23. stated · §2.1 (RoPE applied to first 64 QK dimensions)
  24. stated · §2.2 (learnable attention sink bias s_sink added to the softmax denominator per head, following gpt-oss)
  25. stated · §3.1 Stage 3 (context extended to 256K in the 26-27T range)
  26. stated · Abstract, §1 ('native 32k context length and subsequently extended to 256k')
  27. stated · §3.1, §3.2
  28. stated · Abstract, §3 ('27 trillion tokens')
  29. stated · §3, §3.1
  30. stated · §3.1
  31. stated · §3.2
  32. stated · §2.1, §4.6
  33. stated · §4.1 Stage 1, §4.2 (millions of samples across conversation, reasoning, coding, agent tasks; thinking and non-thinking modes)
  34. stated · §4.4 ('we combine the advantages of MOPD with other types of advantages, such as those computed using Outcome Reward Models (ORMs), including GRPO [shao2024deepseekmath]')
  35. stated · §4.3, Table 8
  36. stated · §4.1 Stage 3, §4.4 (Eq. 5-9)
  37. n/a · modality.type (text-only)
  38. stated · §1, §2.1
  39. stated · §4.6 ('We adopt FP8 for both training and inference')
  40. stated · §2.3, Abstract, §5
  41. stated · report is silent on serving optimizations beyond MTP-based speculative decoding