AI BriefingEgzakta releases Kvark Sparkle 1.1, a Qwen3.8-27B fine-tune for agentic document QA
·2026.11.11 00:00
Key point
Egzakta releases Kvark Sparkle 1.1, an Apache 2.0 fine-tune of Qwen3.8-27B that correctly declines answers not found in documents 91.5% of the time.
Details
- Egzakta released Kvark Sparkle 1.1, an open-weight Apache 2.0 fine-tune of Qwen3.8-27B optimized for agentic document QA with search and read_page tools.
- The model correctly declines to answer when facts are missing from documents 91.5% of the time, compared to 36% for the previous Sparkle 1.0 version and 74% for Sparkle 1.0 with a stop rule.
- Reasoning token usage drops significantly across benchmarks, with reductions of 42% on GPQA Diamond and 48% on FRAMES relative to the base model; on AIME 2026, median tokens dropped from 14,822 to 8,732.
- Performance on SWE-bench Verified is significantly lower than the base model (73.0 vs 81.8), indicating that shorter thinking chains reduce accuracy in agentic coding tasks.
- Available in BF16, FP8, and GGUF formats on Hugging Face, with a recommended vLLM serving configuration using MTP speculative decoding.
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.
Related
Nous Research releases Arete-27B, a post-trained Qwen3.8-27B with reduced state-aligned biasEdge0-145B: 145B-Parameter Open-Source Model Runs on Laptops at 80+ Tokens/s