MiMo-V2.6-Flash-RL
Xiaomi · MiMo-V2 · open weights
- GQA
- Sliding window
- RoPE
- Partial RoPE
- Pre-norm
- MoE
- Dense first layers
- MTP
Facts and where they come from
| Released | 2026-09 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | mit | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 309B | labmodel cardREADME: 309B total / 15B activated parameters |
| Active parameters | 15B | labmodel cardREADME: 15B activated |
| Context length | 1M tokens | labmodel cardREADME: 1M tokens |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for mimo_v2: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for mimo_v2: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | no | codemodelling codeno q/k normalisation in the attention block |
| Positional encoding | RoPE on 33.4% of each head | config.jsonconfig.jsonrope_parameters.partial_rotary_factor |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for mimo_v2: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
9× GQA 64q/4kv + 39× GQA 64q/8kv, window 128. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 309B |
|---|---|
| Active per token (modelled) | 15.4B |
| Without embeddings and output head | 308B total, 14.2B active |
| Multi-token-prediction layers (extra) | 988M |
| Published weights (Hugging Face count) | 311B (packed low-bit tensors, so not comparable) |
| KV cache per token, BF16 (layers that grow with context) | 22.5 KiB |
| KV cache + state at 1M tokens, BF16 | 22.5 GiB |
| Decode FLOPs per token at 4K context | 31.4 GFLOP |
| Prefill FLOPs for a 4K prompt | 120 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 152,576 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 64 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 4 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 192 | config.jsonconfig.jsonhead_dim |
| mixers.full.v_head_dim | 128 | config.jsonconfig.jsonv_head_dim |
| mixers.sliding.type | attn | codemodelling codeattention block |
| mixers.sliding.heads | 64 | config.jsonconfig.jsonswa_num_attention_heads |
| mixers.sliding.kv_heads | 8 | config.jsonconfig.jsonswa_num_key_value_heads |
| mixers.sliding.head_dim | 192 | config.jsonconfig.jsonswa_head_dim |
| mixers.sliding.window | 128 | config.jsonconfig.jsonsliding_window |
| mixers.sliding.v_head_dim | 128 | config.jsonconfig.jsonswa_v_head_dim |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 16,384 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 256 | config.jsonconfig.jsonn_routed_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 2,048 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| layout | 1× full/dense · 4× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe · 5× sliding/moe · 1× full/moe | config.jsonconfig.jsonhybrid_layer_pattern (0 = full, 1 = sliding), moe_layer_freq |
| mtp_layers | 3 | config.jsonconfig.jsonnum_nextn_predict_layers |
| mtp_layer | {"mixer":"sliding","ffn":"dense","n":1} | codemodelling codeMTP block: sliding-window attention + dense MLP |
Sources
- config.json @ 5711b26
- model card
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “Xiaomi MiMo-V2.6-Flash-RL (309B)” (name only; see about).