MiMo-V2-Flash-Base
Xiaomi · base checkpoint
Open in the comparison table Hugging Face ↗ Paper ↗
KV calculations use one sequence at 131,072 tokens and 16-bit cache elements. They include bounded sliding-window caches. This is retained cache payload, not peak prefill memory. The comparison table also offers sourced mixed-precision layouts, including quantization scales.
Release
- Released
2025-12-16
Source: blog · Introducing MiMo-V2-Flash; header
December 16, 2025 Today, we are releasing and open-sourcing MiMo-V2-Flash.
Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.
- License
mit
Source: model-card · README.md; front matter
license: mit
License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.
- Input modalities
text
Inferred from the source; see the explanation below.
Source: model-card · README.md; intro
MiMo-V2-Flash is a Mixture-of-Experts (MoE) language model
No explicit modality statement; config.json (MiMoV2FlashForCausalLM) has no vision or audio config.
Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").
Parameters
- Total params
309B parameters
Source: paper · p.10 §3.2
Overall, MiMo-V2-Flash has 309B total parameters, of which 15B are active.
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
15B parameters
Source: paper · p.10 §3.2
Overall, MiMo-V2-Flash has 309B total parameters, of which 15B are active.
Parameters used per token in the forward pass, as reported by the publisher.
- Active non-embedding (stated)
Not disclosed
Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.
- Counted total params
308.8B parameters
Calculated: embedding + output head + all experts + other weights, XiaomiMiMo/MiMo-V2-Flash-Base@e418701 safetensors headers; reproduces stated total (backbone) and active (with output head)
Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.
- Counted embedding + head
1.25B parameters
Calculated: input embedding + output head, XiaomiMiMo/MiMo-V2-Flash-Base@e418701 safetensors headers
Input embedding plus output head (once if tied), counted from the checkpoint.
- Non-emb. params (counted)
307.5B parameters
Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, XiaomiMiMo/MiMo-V2-Flash-Base@e418701 safetensors headers; reproduces stated total (backbone) and active (with output head)
Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.
- Largest layer (counted)
6.538B parameters
Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, XiaomiMiMo/MiMo-V2-Flash-Base@e418701 safetensors headers
Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.
- Active non-emb. (counted)
14.2B parameters
Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, XiaomiMiMo/MiMo-V2-Flash-Base@e418701 safetensors headers; reproduces stated total (backbone) and active (with output head)
Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.
- Counting scope
MTP inclusion not stated; each MTP block is 0.33B.
Source: paper · p.5 §2.1
The number of parameters for each MTP block is only 0.33B.
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
48
Source: config · /num_hidden_layers
Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.
- d_model
4,096
Source: config · /hidden_size
Residual stream width.
- Vocab
152,576
Source: config · /vocab_size
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Source: config · /tie_word_embeddings
Input embedding and output head share weights.
- d_model / layers
85.3333
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
GQA, SWA
Source: paper · p.5 §2.1
The model consists of 39 SWA layers and 9 GA layers. Both SWA and GA utilize Grouped-Query Attention (GQA)
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Layer layout
Global at layers 0, 5, 11, 17, 23, 29, 35, 41, 47; SWA elsewhere (0-based)
Inferred from the source; see the explanation below.
Source: config · config.json; /hybrid_layer_pattern
"hybrid_layer_pattern": [ 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0 ]
Config hybrid_layer_pattern puts global layers at 0, 5, 11, 17, 23, 29, 35, 41, 47: 4 SWA layers before the first hybrid GA, 5 before each later one. /hybrid_layer_pattern: 1 = SWA, 0 = global (MiMoV2DecoderLayer); ordered backbone only.
Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.
- Global-attention layers
9
Source: paper · p.5 §2.1
The model consists of 39 SWA layers and 9 GA layers.
Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.
- SWA : global
4.33333
Calculated: attn.local_layers / attn.global_layers
attn.local_layers divided by attn.global_layers, the sliding-window to global interleave ratio (5 for '5 SWA : 1 global').
- Sliding-window layers
39
Source: paper · p.5 §2.1
The model consists of 39 SWA layers and 9 GA layers.
Layers with a fixed attention window.
- Query heads
64
Source: paper · p.6 §2.1
SWA has 64 query heads and 8 key-value heads, while GA has 64 query heads and 4 key-value heads.
Config num_attention_heads 64.
Query heads in the global-attention layers.
- KV heads
4
Source: paper · p.6 §2.1
SWA has 64 query heads and 8 key-value heads, while GA has 64 query heads and 4 key-value heads.
Config num_key_value_heads 4.
Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.
- Heads × head dim
12,288
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- Q heads / KV head
16
Calculated: heads / kv_heads
attn.heads divided by attn.kv_heads in global layers.
- QK head dim
192
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- V head dim
128
Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).
- Window
128 tokens
Source: paper · p.5 §2.1
The sliding window size 𝑊 used in MiMo-V2-Flash is 128.
Config sliding_window 128 (read by the SWA mask).
Sliding-window size of the local layers.
- SWA KV heads
8
Source: paper · p.6 §2.1
SWA has 64 query heads and 8 key-value heads, while GA has 64 query heads and 4 key-value heads.
Config swa_num_key_value_heads 8.
KV heads in sliding-window layers (stated even when equal to attn.kv_heads).
- SWA query heads
64
Source: config · /swa_num_attention_heads
MiMoV2Attention.__init__ reads swa_* head counts and dims for SWA layers and the unprefixed keys for GA layers.
Query heads in sliding-window layers (stated even when equal to attn.heads).
- SWA QK head dim
192
Source: config · /swa_head_dim
QK dim; V dim is swa_v_head_dim 128.
Query/key head dimension in sliding-window layers. Value head dimensions may differ; see attn.local_v_head_dim.
- SWA value head dim
128
Source: config · /swa_v_head_dim
Value head dimension in sliding-window layers, when separately specified. KV formulas use this value when present, otherwise the explicit global value dimension, or the local QK dimension when no separate value dimension is given.
- Positional encoding
partial RoPE (first 64 of 192 QK dims)
Source: paper · p.6 §2.1
is partially applied to the first 64 dimensions query and key.
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
5,000,000
Source: config · /rope_theta
640,000 during the 32K stages; raised to 5,000,000 for the 256K extension (§3.2). SWA layers use 10,000.
RoPE base frequency of the global layers at release.
- SWA RoPE base
10,000
Source: config · /swa_rope_theta
MiMoV2FlashRotaryEmbedding(is_swa=True) replaces rope_theta with swa_rope_theta.
RoPE base frequency of the sliding-window layers (stated even when equal to attn.rope_theta).
- Rotary fraction
0.334
Source: config · /partial_rotary_factor
Config value; MiMoV2Attention rotates int(192 × 0.334) = 64 dims, i.e. 1/3.
Fraction of each head's QK dimensions that are rotated.
- QK norm
No
Source: code · modeling_mimo_v2_flash.py; MiMoV2Attention.forward
query_states = self.q_proj(hidden_states).view(qk_hidden_shape).transpose(1, 2)
No norm between projection and RoPE in either layer type.
Queries and keys are normalized per head before the dot product.
- Attention scaling
values × 0.707
Source: code · modeling_mimo_v2_flash.py; MiMoV2Attention.forward
value_states = value_states * self.v_scale
v_scale is config attention_value_scale (0.707), applied in all attention layers; not mentioned in the report. QK scaling is the standard head_dim**-0.5.
Query-key scaling when it is not 1/sqrt(head_dim), e.g. query_pre_attn_scalar, length-dependent (SSMax-style) scaling, or value scaling.
- Attention sinks
learnable per-head sink bias in the softmax denominator (gpt-oss style), SWA layers only
Source: paper · p.6 §2.2
our implementation follows the design used in gpt-oss
Config: add_swa_attention_sink_bias true, add_full_attention_sink_bias false.
Learned sink logits or sink bias in the softmax, with the layer types that use them.
- QKV bias
No
Source: config · /attention_bias
Query/key/value projections have bias terms.
Feed-forward
- Activation
SwiGLU
Source: code · modeling_mimo_v2_flash.py; MiMoV2MLP.forward
down_proj = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
act_fn is config hidden_act silu; routed experts use the same MiMoV2MLP.
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
16,384
Source: paper · p.10 §3.2
The intermediate hidden dimension of the FFN of dense layers is set to 16384.
Dense FFN in layer 0 only; config intermediate_size 16384.
Intermediate width of the dense FFN layers that are actually executed.
- Dense FFN / d_model
4
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Mixture of experts
- Routed experts
256
Source: paper · p.10 §3.2
Each MoE layer contains 256 routed experts, with 8 experts activated per token, and an intermediate hidden dimension of 2048 for each expert.
Config n_routed_experts 256.
Routed experts per MoE layer.
- Experts per token
8
Source: paper · p.10 §3.2
Each MoE layer contains 256 routed experts, with 8 experts activated per token, and an intermediate hidden dimension of 2048 for each expert.
Config num_experts_per_tok 8.
Routed experts selected per token.
- Expert dim
2,048
Source: paper · p.10 §3.2
Each MoE layer contains 256 routed experts, with 8 experts activated per token, and an intermediate hidden dimension of 2048 for each expert.
Config moe_intermediate_size 2048.
Intermediate width of one routed expert.
- Leading dense layers
1
Source: paper · p.10 §3.2
All layers except the first are equipped with sparse MoE.
Config moe_layer_freq is 0 for layer 0 only.
Number of initial blocks that use a dense FFN instead of MoE.
- Router score
sigmoid
Source: config · /scoring_func
Gating function over expert logits (softmax, sigmoid).
- Load balancing
aux-loss-free expert bias (update 0.001 in stages 1–2, 1e-5 in stage 3) + sequence aux loss 1e-5
Source: paper · p.10 §3.2
the MoE sequence auxiliary loss coefficient is set to 1.0 × 10−5 for all stages.
Config topk_method noaux_tc: the expert bias is added to scores for selection only.
Load-balancing method, with coefficients where disclosed.
- Renormalize top-k
Yes
Source: config · /norm_topk_prob
Selected expert weights are renormalized to sum to 1.
- Expert dim / d_model
0.5
Calculated: moe.expert_dim / d_model
moe.expert_dim divided by shape.d_model.
- Sparsity (experts / top-k)
32
Calculated: experts / top_k
Routed experts divided by experts selected per token.
- Active / total
0.0485437
Calculated: params.active / params.total
params.active divided by params.total.
Multi-token prediction
- MTP depth
1
Source: paper · p.10 §3.2
The model uses a single MTP layer during pre-training.
Additional future tokens predicted during pretraining (0 when the model was trained without MTP).
- MTP modules
1
Source: paper · p.9 §2.3.2
During pre-training, only a single MTP head is attached to the model to avoid extra training overhead.
Released checkpoint holds three modules: the head replicated and trained jointly in post-training.
MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.
- MTP module
SWA (64 Q / 8 KV heads, sink bias) + dense SwiGLU FFN; RMSNorm'd hidden state and next-token embedding concatenated and linearly projected
Source: paper · p.5 Figure 2 caption; §2.1
The MTP blocks employ SWA and a dense FFN.
Input combination from Figure 2 (two RMSNorms, Linear) and index weights enorm, hnorm, eh_proj; SWA heads from Table 1.
Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).
- MTP FFN
dense
Source: paper · p.9 §2.3.2
We use a small dense FFN rather than MoE to limit parameter count
Width not stated.
FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.
- MTP loss weight
0.3 in stage 1; 0.1 in stages 2 and 3
Source: paper · p.10 §3.2
The MTP loss weight is set to 0.3 for Stage 1 and 0.1 for Stage 2 and 3.
Weight of the MTP loss, including its schedule.
- MTP weights released
Yes
Source: config · model.safetensors.index.json; weight_map
"model.mtp.layers.2.eh_proj.weight"
Three SWA + dense-FFN MTP blocks; the single pretraining head is replicated and trained jointly in post-training (§2.3.2).
The released checkpoint contains the MTP module weights.
Normalization and stability
- Norm
RMSNorm
Source: code · modeling_mimo_v2_flash.py; MiMoV2DecoderLayer.__init__
self.input_layernorm = MiMoV2RMSNorm(config.hidden_size, eps=config.layernorm_epsilon)
Normalization layer type in the blocks.
- Norm placement
pre
Source: code · modeling_mimo_v2_flash.py; MiMoV2DecoderLayer.forward
hidden_states = self.input_layernorm(hidden_states)
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-05
Source: config · /layernorm_epsilon
Epsilon of the block normalization layers.
Context length
- Pretraining context
32.77K tokens
Source: paper · p.10 §3.2
we set the pre-training sequence length to 32,768
Sequence length of the main pretraining stage.
- Trained context
262.1K tokens
Source: paper · p.10 §3.2
the sequence length is extended to 262,144
Longest sequence length trained on after context extension.
- Advertised context
262.1K tokens
Source: model-card · README.md; §1 Introduction
The context window supports up to 256k length.
Config max_position_embeddings 262144.
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
- Extension method
Stage 3 (26T–27T tokens): 32K → 256K, GA RoPE base 640K → 5M
Source: paper · p.10 §3.2
In Stage 3, the sequence length is extended to 262,144, and the RoPE base frequency of GA is adjusted to 5,000,000.
How context was extended (stages, RoPE base change, YaRN factor).
Pretraining recipe
- Pretraining tokens
27T tokens
Source: paper · p.1 Abstract
The model is pre-trained on 27 trillion tokens with Multi-Token Prediction (MTP)
Includes mid-training (stage 2) and context extension (stage 3).
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
AdamW
Source: paper · p.10 §3.2
We employ the AdamW optimizer with 𝛽1 = 0.9, 𝛽2 = 0.95, and a weight decay of 0.1.
Optimizer used for the bulk of parameters.
- Peak LR
0.00032
Source: paper · p.10 §3.2
the learning rate starts with a linear warm-up from 0 to 3.2 × 10−4 over the first 50B tokens
Peak learning rate of the main pretraining stage.
- Final LR
1e-05
Source: paper · p.10 §3.2
The learning rate in Stage 3 decays from 3.0 × 10−5 to 1.0 × 10−5 following a cosine schedule
Learning rate at the end of the decay.
- LR schedule
Stage 1: 50B warmup, constant 3.2e-4 for 12T, cosine to 1e-4 over 10T; stage 2: cosine 1e-4 → 3e-5 over 4T; stage 3: cosine 3e-5 → 1e-5
Source: paper · p.10 §3.2
followed by a constant phase at 3.2 × 10−4 for 12T tokens, and concludes with a cosine decay to 1.0 × 10−4 over 10T tokens.
Learning-rate schedule shape with phase boundaries.
- Warmup
50B tokens
Source: paper · p.10 §3.2
the learning rate starts with a linear warm-up from 0 to 3.2 × 10−4 over the first 50B tokens
Warmup length in steps or tokens.
- Batch size
2,048 × 32,768 (67.1M tokens) after a linear warmup over the first 500B tokens; 256 × 262,144 (67.1M tokens) in stage 3
Source: paper · p.10 §3.2
The batch size warms up linearly to 2,048 over the initial 500B tokens and remains constant for the remainder of both stages.
The report does not give the unit; read as sequences because both stages then give the same tokens per step.
Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.
- Batch (tokens)
67.11M tokens
Inferred from the source; see the explanation below.
Source: paper · p.10 §3.2
The batch size warms up linearly to 2,048 over the initial 500B tokens
Unit not stated; read as sequences: 2,048 × 32,768. Stage 3's 256 × 262,144 gives the same total.
Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).
- Weight decay
0.1
Source: paper · p.10 §3.2
We employ the AdamW optimizer with 𝛽1 = 0.9, 𝛽2 = 0.95, and a weight decay of 0.1.
Decoupled weight decay coefficient.
- Adam betas
(0.9, 0.95)
Source: paper · p.10 §3.2
We employ the AdamW optimizer with 𝛽1 = 0.9, 𝛽2 = 0.95, and a weight decay of 0.1.
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
1
Source: paper · p.10 §3.2
Gradient clipping is applied with a maximum norm of 1.0.
Global gradient-norm clipping threshold.
- Training precision
FP8 mixed precision (DeepSeek-V3-style); BF16 attention output projections, embedding and output head; FP32 router
Source: paper · p.6 §2.1
we retain BF16 precision for the attention output projections, as well as for the embedding and output head parameters, while maintaining FP32 precision for the MoE router parameters.
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Initialization
std 0.006 for all parameters
Source: paper · p.10 §3.2
All learnable parameters are randomly initialized with a standard deviation of 0.006.
Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.
- Data mix
web, books, papers, code, math, STEM; upweighted long-form web documents and repository-level code, PRs, issues and commits; ~5% synthetic reasoning data from stage 2
Source: paper · pp.9–10 §3, §3.1
we emphasize long-form web documents and carefully curated code corpora
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Stages
pre-training
- Tokens
- 22000000000000
- Seq len
- 32768
- Lr
- 0.00032
- Batch sequences
- 2048
Source: paper · p.10 §3.1, §3.2
Stage 1 (Pre-training, 0 – 22T).
General corpus; warmup, constant, then cosine to 1e-4.
mid-training
- Tokens
- 4000000000000
Source: paper · p.10 §3.1, §3.2
Stage 2 (Mid-training, 22 – 26T).
Code upsampled, ~5% synthetic reasoning data; LR cosine 1e-4 → 3e-5.
context-extension
- Tokens
- 1000000000000
- Seq len
- 262144
- Batch sequences
- 256
Source: paper · p.10 §3.1, §3.2
Stage 3 (Context Extension, 26 – 27T).
Stage 2 mix with long-range-dependency data upsampled; LR cosine 3e-5 → 1e-5.
Source: paper · p.10 §3.1, §3.2
Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.
Scaling & HP transfer
- Scaling ladder
attention ablations on a 32B dense model: 250B tokens at 8K, then 40B at 32K
Source: paper · p.6–7 §2.2.1
we conduct exploratory and empirical studies on a 32B dense model
Token budget from §2.2.1 Baselines and Benchmarks.
Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.
- Tokens / param
87.3786
Calculated: train.tokens / params.total
train.tokens divided by params.total.
- Tokens / active param (non-embedding)
1,901.94
Calculated: train.tokens / params.counted_active_non_embedding
train.tokens divided by params.counted_active_non_embedding, or by the stated params.active_non_embedding when there is no validated checkpoint count.
- AdamW timescale τ
0.0776723
Calculated: train.batch_tokens / (train.lr_peak × train.weight_decay × train.tokens)
B / (η λ D) with B = train.batch_tokens, η = train.lr_peak, λ = train.weight_decay and D = train.tokens: the AdamW EMA timescale as a fraction of training (Bergsma et al. 2025, Power Lines, arXiv 2505.13738). Computed only when the optimizer is AdamW; uses the peak LR.
KV cache
- KV bytes / token
23,235 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 9 × (1280 × 16/8) B/entry; SWA KV: 39 × (2560 × 16/8) B/entry
Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.
- Global KV bytes / token
23,040 bytes
Calculated: 23040 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 9 × (1280 × 16/8) B/entry; SWA KV: 39 × (2560 × 16/8) B/entry
Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.
- KV bytes / token / B active
1,549 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 9 × (1280 × 16/8) B/entry; SWA KV: 39 × (2560 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
203,030,528 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 9 × (1280 × 16/8) B/entry; SWA KV: 39 × (2560 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
3,045,457,920 bytes
Calculated: 23040 B × 131072 (Global KV) + 199680 B × min(131072, 128) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 9 × (1280 × 16/8) B/entry; SWA KV: 39 × (2560 × 16/8) B/entry
Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.
Evaluations
- AA Intelligence Index
21
Source: leaderboard · Artificial Analysis model page /models/mimo-v2-flash-reasoning, Wayback snapshot 2026-10-19
MiMo-V2-Flash (Reasoning) scores 21 on the Artificial Analysis Intelligence Index
Score of MiMo-V2-Flash, post-trained from this base. Variant: MiMo-V2-Flash (Reasoning). AA entry for the December 2025 release. Index v4.3.2.
Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).