Strata routing test: route prior boosts decode 45% with minimal quality loss
blakeman8192 · reddit · 2026-10-29
A Reddit user benchmarked Strata v0.2.7.1's two routing optimizations on Qwen 3.8 Flash Next IQ3S across code, math, and instruction-following evals.
- STRATAROUTEPRIOR adds a 1.5 bonus (4.48x multiplier) to experts already cached in VRAM, letting cached experts replace close-scoring uncached ones.
- STRATAROUTETAILSKIP (default 7) simply drops contributions from ranked experts 7-10 if not in VRAM.
Results: both off = 72.3 tok/s, 20% code, 76.7% AIME, 81.3% IFEval. Route prior alone = 104.6 tok/s (+45%), 16.7% code, 66.7% AIME, 78.7% IFEval. Tail skip alone = 83.9 tok/s but AIME dropped to 63.3%.
A second round with larger thinking budgets showed route prior's quality gap mostly disappears with longer reasoning (80% AIME), while enabling both settings performed worst (71.8% IFEval).
Verdict: enable route prior for everything but the hardest tasks; tail skip degrades more than it's worth.
More from Infra
- RTX 5080 now nears $2,000 at US retailers, up $800 in two months — chemist_slime · 2026-10-29
- Mac Studio M5 Ultra with 512GB memory starts at $22,299 — Boomfrag · 2026-10-29
- 230 agent-run experiments on LLM inference: 8 lessons from DeepSeek 4.1 Flash tuning — jiayq · 2026-10-29
- Together's Post-Training Uses Per-GPU Pricing to Cut Agent Rollout Costs — togethercompute · 2026-10-29
- GitHub Unveils New Git Architecture Hitting 7,200 Writes/sec for AI Collaboration — mariorod1 · 2026-10-29
- Running a 10B-parameter model on a garbage GPU is the late-night thrill of choice — lxfater · 2026-10-29