Research
We make inference cheaper and faster, and measure every result in production.
Articles
Synth: Better Answers From Documents You Already Retrieved
Sparset’s open-weight Qwen3.5-4B fine-tune for document synthesis scored 95.22 vs 91.16 on an internal 500-case test. See the recipe, results and limits.
Nanolite Models: Cutting The Reasoning Models Don’t Need
We explore how fine-tuning changes generation in two small reasoning models. Nanolite 0.6B delivered answers 3.28× faster in a workplace pilot, while Nanolite V2 used about 92% fewer tokens on mutually completed GSM8K and BBH questions. The experiments reveal both the potential of shorter reasoning and the quality tradeoffs behind the speedups.
Subset: turning LLMs into 26x faster decision-making machines
An experimental engine that reads a pretrained LLM’s scores for known answers instead of generating JSON. 26x faster by median latency, limits stated.