Mixtral 8x22B
Mistral AI · Mixture-of-Experts (MoE) · License: Apache-2.0
Mistral AI's large sparse MoE model, combining eight 22B expert networks. One of the earliest fully-permissive-licensed models at this scale.
Memory requirement — the visual calculator
In real deployments, memory needs also depend on numeric precision (e.g. FP16 vs INT8 vs INT4), which changes bytes-per-parameter. This store uses a simplified, transparent formula so beginners can follow the logic: (parameters in billions × 1 GB) + 20% overhead for the runtime, active requests, and context memory.
Calculation: 141B parameters × 1 GB = 141 GB, plus 20% (28.2 GB) = 169.2 GB total The dedicated memory built into the graphics card. An AI model's entire set of "weights" (its learned parameters) must fit into this memory before the card can run it. If a model needs more VRAM than one card has, you need multiple cards working together, or a card with more memory. needed. Note: even though only 39B parameters ( A model design where only a fraction of the model's parameters ("active parameters") are used to process any single piece of text, even though the full model is much larger. This makes the model faster to run than its total size suggests, but you still need enough memory to hold the entire model, not just the active part. ) are active per token, the full 141B must still be loaded into memory.
Leaderboard
Debuted on Chatbot Arena (LMSYS) around rank #11-13 in April 2024.