Egzakta AI Lab open-sources KVARK Sparkle 1.0 model

On 2026-11-02 at 21:48, Egzakta AI Lab announced its first open-weight model, KVARK Sparkle 1.0 (27B), in a post on the Reddit community LocalLLaMA. The post, written by someone identifying as part of the team behind KVARK, says the model is built on Qwen3.8-27B and focuses on making reasoning more efficient: fewer thinking tokens, faster responses, and comparable answer quality. The summary accompanying the release gives a concrete figure, stating that KVARK Sparkle 1.0 can cut thinking tokens by up to 67% while keeping answer quality at the same level and speeding up responses. The team explains the motivation as follows: reasoning models can burn a lot of tokens on tasks that do not actually require that much reasoning, and the goal was to see how much of that overhead could be removed. That sentence is cut off mid-thought in the source, so the full reasoning and any further claims were not disclosed. As of now, the public record contains only the model's positioning and efficiency claims; no details on download channels, license terms, or benchmark results have appeared, and there has been no follow-up post since the initial release announcement.