Strata inference engine silently caps thinking at 75% with an injected wrap-up line
Regular-Register-358 · reddit · 2026-11-03
A Reddit user found that the Strata/PolyStrata local inference engine (38k GitHub stars) has a default-on "answerroom" behavior: with thinking on, maxtokens ≥ 1024, and no explicit reasoningbudgettokens, it silently caps thinking at 0.75 × maxtokens by injecting a first-person wrap-up line — "I have thought about this long enough; time to give my answer" — plus </think> into the model's inner monologue, then continues the same reply. The client gets no error, flag, or finish reason, and often no log. The post links exact code locations (wrap-up message L225, 3/4 trigger L5491, injection point L5848, default-on L4181) at commit c62d5e3. Implication: benchmarks comparing Strata-served models against other engines may involve mid-thinking nudges that are hard to attribute as helpful or harmful.
More from Infra
- Open-model CT and malaria demos run on a single GPU — where should hospitals put it? — MaziyarPanahi · 2026-11-03
- Benchmarking System 1 decision models: 40 yes/no questions in one pass, 159ms agent routing — ThatOneSnare · 2026-11-03
- Could a Folding@home-style P2P network of idle GPUs collectively run and train LLMs? — lenicalicious · 2026-11-03
- ISTA (GSQ-RCO) vs Unsloth quant beef heats up as users ask why ISTA won't publish KLD — Queasy_Back2588 · 2026-11-03
- Has anyone benchmarked int8 models before and after the RTX 20-40 series unlock? — Dave-C · 2026-11-03
- B300/GB300 Blackwell financing deals smell like bogus multi-layer SPV schemes, warns insider — TheZachMueller · 2026-11-03