Active Parameters
1.02T
Context Length
1.05M
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
420 KB/tok
License
MIT License
Release Date
21 Sept 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Self-hosted only
Precision
Context
Workload Memory: 1983.91 GB
FP16 · 8K context
Base Weights: 1978.13 GB
KV Cache: 3.28 GB
Runtime & Graphs: 2.50 GB
16x NVIDIA H200 141GB
1983.9 / 2030.4 GB usable · TP 8 · PP 2
Determine requirements for your specific workload:
Upgrade to Pro
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 5.3.0
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
No evaluation benchmarks for MiMo V2.6 Pro RL available.
Overall Rank
-
Coding Rank
-
MiMo V2.6 Pro RL is Xiaomi's flagship reasoning-oriented model trained using reinforcement learning for complex problem solving and agent workflows. It achieves enhanced performance across coding, mathematics, and complex multi-step tasks.
Attention
Attention Structure
Single-Head Attention
Attention Heads
128
Key-Value Heads
8
Attention Head Dimension
192
Position Embedding
ROPE
RoPE Theta
10,000,000
Sliding Window Attention
Yes
Sliding Window Size
128
Sliding Window Ratio
85.7%
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
6,144
Number of Layers
70
FFN Intermediate Size (Dense)
16,384
Multi-Token Prediction Heads
5
Tokenizer
Vocabulary Size
152,576
Mixture of Experts
Total Expert Parameters
42.0B
Number of Experts
384
Active Experts
8
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
1
MiMo-V2-Flash is a Mixture-of-Experts (MoE) model with hybrid attention architecture designed for high-speed reasoning and agentic workflows. It features Multi-Token Prediction (MTP) to achieve state-of-the-art performance while significantly reducing inference costs. The model is optimized for long-context modeling and efficient inference.