Show HN: Kvark Sparkle 27B – reasoning model trained to think with fewer tokens
Category: ai-ml
Tags: reasoning-model, fine-tuning, grpo, speculative-decoding, qwen, llm, lora
Score: 6.0/10 (Innovation: 5, Technical: 7, Documentation: 8, Utility: 6)
Kvark Sparkle 1.0 is a 27B reasoning model fine-tuned from Qwen3.8-27B that produces correct answers with roughly 60% fewer reasoning tokens, using a two-line LoRA training recipe (efficiency GRPO + knowledge SFT) merged via weight averaging and GSPO. It claims near-parity accuracy with its base model and significant speed gains on math/code benchmarks, with ready-to-run FP8/GGUF variants for vLLM and llama.cpp.
Target audience: AI/ML researchers and engineers deploying reasoning LLMs
Repository: https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B
View on Hacker News