A mixture-of-experts (MoE) model divides parameters among specialists and activates only a few for each token. This reduces per-token computation without reducing the memory required to keep the full model resident: an 80B model with 3B active parameters can approach 3B-model speed while consuming 80B-model memory. section 9.3.4
Local MoE deployment is governed by hardware thresholds:
- Below 16GB, local models mainly support tab completion, boilerplate, imports, and pattern continuation. section 9.3.4
- At roughly 24–32GB, tab completion, inline edits, and single-file work become useful, while multi-file reasoning remains unreliable. section 9.3.4
- At 48GB or more, large MoE models can support multi-file agent work, but long contexts add KV-cache memory costs. section 9.3.4
A large model with small active compute
Qwen3-Coder-Next is described as an 80B MoE model that activates 3B parameters per token, achieves genuine multi-file capability on a 48GB unified-memory Mac, and still trails frontier models on the hardest cross-cutting tasks. section 9.3.1









