MiMo-V2.6-Pro-RL

ADDED 2026-10-18. Open-weight sparse MoE, the flagship checkpoint of the MiMo-V2.6 series and the larger sibling of models/mimo-v2-6-flash-rl. Native omnimodal (text, image, video, audio) with a 5-layer SWA speculative decoder the card describes as DFlash-style, predicting 7 tokens per forward pass. RUNNABLE WITH EXPERT OFFLOAD: the card's Model Summary states 1.02T total and 42B activated, and the architecture table confirms 384 routed experts with 8 activated and no shared expert. active_params_b is that 42B vendor figure, not a count derived here. Only layer 0 is a dense FFN (moe_layer_freq is 0 followed by 69 ones). Eight experts per token is the same routing width as models/mimo-v2-6-flash-rl and models/qwen3-6-35b-a3b, so the per-token expert traffic stays a small slice of the checkpoint and the model stays fast with experts in RAM. A dense model at this size would decode at RAM bandwidth. params_b stays 1000: the HF model page reports '1T params' and the card rounds the total to 1.02T, which is the same figure at the precision this record stores. flops_per_token_active = 2 x 42e9 = 8.4e10 is the schema one-pass approximation from the card's 42B activated figure. GEOMETRY from config.json fetched 2026-10-18: hidden 6144, 70 layers, global attention 128 query heads / 8 kv heads / head_dim 192 / v_head_dim 128 so gqa_ratio 16, and the sliding layers match it (swa_num_attention_heads 128, swa_num_key_value_heads 8, swa_head_dim 192, swa_v_head_dim 128, sliding_window 128). hybrid_layer_pattern has 10 global (value 0) and 60 sliding (value 1); global layers sit at indices 0, 7, 15, 23, 31, 39, 47, 55, 62, 69, which is the card's 70 / 60 / 10 split. MoE: n_routed_experts 384, num_experts_per_tok 8, n_shared_experts null so shared_expert is false, moe_intermediate_size 2048. silu, vocab 152576, tie_word_embeddings false, partial_rotary_factor 0.334, rope_theta 10000000, rope_type default, no rope_scaling key, so context_scaling_method is native. The checkpoint is FP8 (quantization_config fmt e4m3, activation_scheme dynamic). KV (bf16), growing term only: the 10 global layers cache K at head_dim 192 and V at v_head_dim 128, so 10 x 8 kv heads x (192 + 128) x 2 bytes = 51,200 B/token. The 60 sliding layers hold a fixed 128-token window at the same per-head geometry (8 kv heads, K 192 + V 128), 60 x 8 x 320 x 2 x 128 = 39.3 MB that does not grow with context, so kv_cache_bytes_per_token records the unbounded term only. If all 70 layers cached the full context it would be 358,400 B/token, about 7x the recorded number. BENCHMARKS ARE VENDOR CLAIMS. BEATS models/mimo-v2-6-flash-rl, the catalog's best published Terminal-Bench 2.1 number, on Terminal-Bench 2.1: 89.9 vs 87.6. Both numbers are vendor claims from the same Xiaomi table (the card's 'Terminal Bench 2.1' column), and the HF eval-results widget repeats 89.9 on harborframework/terminal-bench-2.1 with verified=false. The same table gives Terminal Bench 4.0 34.9, DeepSWE v1.1 71.9 and Toolathlon-Verified 76.9. The card publishes no SWE-bench Verified or SWE-bench Pro number, so those are not claimed here.