Individual Runs (2026-07-06)
Seven models/configurations tested using llama-benchy v0.4.0 on the same stack.
| Model | Config | Key result |
|---|---|---|
| Ornith-IQ4 (IQ4_XS-MTP-graft) | IQ4_XS + Q6_K head | −18% smaller, +10% faster, better KLD — new prod candidate |
| Ornith-1.0-35B (prod) | Q4_K_M + MTP n2 | Flat ~76 t/s 0→16k, shallow −24% decay to 65k |
| Qwen3.6-35B-A3B vocab-trimmed | Q4_K_S + MTP n2 | Shallowest decay (−19%), pulls ahead at deep context |
| Qwen3.6-35B-A3B MTP OFF vs ON | Q4_K_M | Cleanest isolation: MTP = +58% at 65k |
| Qwen3.6-35B-A3B DFlash vs MTP | Q4_K_M | MTP wins (92% accept vs ~60%), DFlash stable with -ub 512 |
| gemma-4-12b-it (dense) | Q4_K_M + MTP n3 | Dense penalty: ~2× slower than A3B models |
| Qwen3.6-27B (dense) | Q4_K_M + MTP n2 | Slowest model (23 t/s); MTP net-neutral on dense |