absent = the key is not in that model’s config.json (the model then uses its library default).

Changed fields

field MiMo-V2.6-Pro-RL MiMo-V2.6-Flash-RL
bytes 566027990784 172923364096
hidden_size 6144 4096
hybrid_layer_pattern 10 × 0, 60 × 1 9 × 0, 39 × 1
moe_layer_freq 1 × 0, 69 × 1 1 × 0, 47 × 1
n_routed_experts 384 256
num_attention_heads 128 64
num_hidden_layers 70 48
num_key_value_heads 8 4
num_nextn_predict_layers absent 3
swa_num_attention_heads 128 64

Key unchanged fields

field value
head_dim 192
max_position_embeddings 1048576
moe_intermediate_size 2048
n_shared_experts null
num_experts_per_tok 8
rope_theta 10000000
sliding_window 128
swa_head_dim 192
swa_num_key_value_heads 8
swa_rope_theta 10000
swa_v_head_dim 128
tie_word_embeddings false
v_head_dim 128
vocab_size 152576

Other differing config keys

field MiMo-V2.6-Pro-RL MiMo-V2.6-Flash-RL
attention_value_scale 0.612 0.707
audio_config.out_hidden_size 6144 4096
layernorm_epsilon 1e-05 1e-06
vision_config.out_hidden_size 6144 4096

Note: vision_config and audio_config present in config.json – the stored-element and byte counts below include those encoders, not the text model alone.

Dtype / quantization notes

MiMo-V2.6-Pro-RL:

  • “activation_scheme”: “dynamic”
  • “fmt”: “e4m3”
  • “mxfp4_block_size”: 32
  • “quant_method”: “fp8”
  • “store_dtype”: “mxfp4”
  • computed: dtype BF16: 10,647,286,656 stored elements, 21,294,573,312 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype F32: 857,088 stored elements, 3,428,352 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype F8_E4M3: 13,378,781,184 stored elements, 13,378,781,184 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype U8: 531,351,207,936 stored elements, 531,351,207,936 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: sum of prod(shape) over 160,040 tensors in 130 safetensors shard(s) = 555,378,132,864 stored elements. Not a parameter count: config.json declares a quantization_config; 79488 weight tensor(s) are U8/I8 with a matching *_scale companion (packed storage); config.json declares a vision_config; config.json declares an audio_config.

MiMo-V2.6-Flash-RL:

  • “activation_scheme”: “dynamic”
  • “fmt”: “e4m3”
  • “mxfp4_block_size”: 32
  • “quant_method”: “fp8”
  • “store_dtype”: “mxfp4”
  • computed: dtype BF16: 4,101,308,032 stored elements, 8,202,616,064 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype F32: 248,192 stored elements, 992,768 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype F8_E4M3: 3,859,808,256 stored elements, 3,859,808,256 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: dtype U8: 160,859,947,008 stored elements, 160,859,947,008 bytes (from each tensor’s data_offsets span in the safetensors header)
  • computed: sum of prod(shape) over 73,081 tensors in 65 safetensors shard(s) = 168,821,311,488 stored elements. Not a parameter count: config.json declares a quantization_config; 36096 weight tensor(s) are U8/I8 with a matching *_scale companion (packed storage); config.json declares MTP draft layers; config.json declares a vision_config; config.json declares an audio_config.

Sources