Qwen3 30B A3B
30B of knowledge, 3B of work per word
Install with Bobbin
Don't have the app? Download it here.
A mixture-of-experts model that holds 30B parameters but activates only about 3B for each word. That is what makes near-flagship quality possible on a consumer card at usable speed.
This is the model that makes the case for running AI locally at all.
Available versions
| Quantization | Size | VRAM | RAM | Context | |
|---|---|---|---|---|---|
| Q4_K_M | 18.6 GB | 18.0 GBest. | 27.8 GB | 8k | Install |
Memory needed grows with the length of a conversation, so these figures are for the context shown. Anything marked est. is worked out from the file size rather than measured on real hardware — the app says so there too. The desktop app checks all of it against your actual machine before installing.