Kvark Sparkle 1.0: Qwen3.8-27B Fine-Tune Cuts Reasoning Tokens by 60-67% with Apache 2.0 License

·2026.10.30 00:50

Key point

The Apache 2.0 release maintains base accuracy within confidence intervals while reducing median wall-clock time from 28.1s to 13.7s on a single H200.

Details

Fine-tuned Qwen3.8-27B model named Kvark Sparkle 1.0 reduces median reasoning tokens by 60-67% compared to the base model.
Median wall-clock time drops from 28.1s to 13.7s on a single H200 using vLLM 0.30 and FP8, with decode speed remaining consistent at 165 vs 171 tok/s.
Accuracy matches the base model within 95% bootstrap confidence intervals, with GPQA Diamond at 89.3% and AIME at 93.3%.
Training uses SFT on shortest correct traces and GSPO with accuracy rewards, avoiding proprietary model outputs.
MTP head functionality is preserved, allowing speculative decoding with 3 speculative tokens.
Available in BF16, FP8, and GGUF formats on Hugging Face under the Egzakta organization.

open-source vllm fine-tuning inference-optimization apache-2.0 qwen3.8-27b kvark-sparkle-1.0

Read originalShare on Threads

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.