The lab
Notes, published with their data.
Aelius runs Qwen3.6-27B, an open-weights hybrid-attention model, on a single machine we own. These notes document how — including the configurations that failed, the measurements that decided, and the traps that cost us time. Everything here is reproducible on consumer hardware.
Serving 84K context on a 24 GB card
The context ceiling wasn't VRAM — it was a 1 GiB safety margin inside llama.cpp's auto-fit. Finding it bought 31% more context and gave up nothing.
Read the note →Speculation has a depth limit
Multi-token prediction at depth 2 runs 65% faster than depth 5 — deeper drafts evict a layer to CPU and halve their own acceptance rate. Measured, with seeds.
Read the note →The economics of hybrid attention
Only 16 of Qwen3.6's 64 blocks pay per-token memory. That single design choice is why a 200-page conversation fits on one consumer GPU — and why Aelius can be free.
Read the note →