MiMo-V2-Flash

View as Markdown

MiMo-V2-Flash is Xiaomi’s hybrid attention Mixture-of-Experts language model. It alternates full and sliding-window attention layers, uses a sigmoid_with_bias router with group-limited expert routing, and ships as an FP8 HF checkpoint.

TaskText Generation (MoE, hybrid attention)
ArchitectureMiMoV2FlashForCausalLM
ParametersApprox. several hundred B total / much smaller active
HF OrgXiaomiMiMo

Available Models

  • MiMo-V2-Flash: hybrid full/sliding-window attention with FP8 weights.

Architecture

  • MiMoV2FlashForCausalLM
  • Sliding-window attention via the MiMoV2FlashAttention(is_swa=True) path.
  • MoE blocks use the shared nemo_automodel.components.moe.layers.MoE with score_func="sigmoid_with_bias" and gate_precision=fp32 so routing decisions stay numerically stable when activations are bf16.
  • FP8 round-trip in MiMoV2FlashStateDictAdapter covers the bulk of attention/expert weights; layer norms, the gate, lm_head, and embed_tokens stay in bf16 per NON_QUANTIZED_KEY_PATTERNS.

Example HF Models

ModelHF ID
MiMo-V2-FlashXiaomiMiMo/MiMo-V2-Flash

Example Recipes

RecipeDescription
mimo_v2_flash_hellaswag.yamlSFT — MiMo-V2-Flash on HellaSwag

Try with NeMo AutoModel

1. Install (full instructions):

$pip install nemo-automodel

2. Clone the repo to get the example recipes:

$git clone https://github.com/NVIDIA-NeMo/Automodel.git
$cd Automodel

3. Run the recipe from inside the repo:

$automodel --nproc-per-node=8 examples/llm_finetune/mimo_v2_flash/mimo_v2_flash_hellaswag.yaml

1. Pull the container and mount a checkpoint directory:

$docker run --gpus all -it --rm \
> --shm-size=8g \
> -v $(pwd)/checkpoints:/opt/Automodel/checkpoints \
> nvcr.io/nvidia/nemo-automodel:26.06.00

2. Navigate to the AutoModel directory:

$cd /opt/Automodel

3. Run the recipe:

$automodel --nproc-per-node=8 examples/llm_finetune/mimo_v2_flash/mimo_v2_flash_hellaswag.yaml

See the Installation Guide and LLM Fine-Tuning Guide.

Fine-Tuning

See the LLM Fine-Tuning Guide.

Hugging Face Model Cards