Egzakta AI Lab releases KVARK Sparkle 1.0, a Qwen3.8-27B derivative with up to 67% fewer thinking tokens

·2026.11.02 22:48

Key point

The open-weight model achieves a 13.7s median response time versus 28.1s for the base model while maintaining comparable accuracy on five benchmarks.

Egzakta AI Lab released KVARK Sparkle 1.0 (27B), an open-weight model based on Qwen3.8-27B designed to reduce reasoning overhead.
The model achieves up to 67% fewer thinking tokens and a 2x faster median response time (13.7s vs. 28.1s) compared to the original Qwen3.8-27B.
Accuracy remains comparable across GPQA Diamond, AIME, LiveCodeBench, IFEval, and MMLU-Pro, with no statistically significant differences in testing.
Practical tasks show speedups of 1.6x for dashboard generation, 1.4x for landing page generation, and 3.2x for SVG illustration.
Weights are available under Apache 2.0 in BF16, FP8, and GGUF formats, supporting llama.cpp, vLLM, Transformers, and SGLang.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.