- Long agent conversations no longer get refused early: a 27B conversation on a 24 GB budget now runs to 13,274 tokens instead of 9,475 (#1300).
- 27B agent turns are faster: a cached turn on an M5 Pro takes 151-173 ms instead of 208-301 (#1276), and a 4K prompt on a Mac mini M6 prefills at 589 tok/s instead of 492 (#1275).
- Fixes garbage output from Nemotron 3.5 Super VL on M5 Ultra (#1301), and top_k 0 or above 64 is now honored instead of silently clamped (#1286).