KVARK Sparkle 1.1 (27B, Apache 2.0): Refuses When Docs Lack the Answer, Uses 7-48% Fewer Tokens
dzulajla98 · reddit · 2026-11-10
The KVARK team released Sparkle 1.1 (27B, Apache 2.0, with MTP head), targeting document Q&A: it searches a document base, answers from what it finds, and declines when the answer isn't there. Over 1.0: doc Q&A score 89.5→91.0; correct refusals for unanswerable questions jumped from 74% to 91.5%. It uses 7-48% fewer reasoning tokens than Qwen3.8-27B (median) with no statistically significant accuracy difference on GPQA Diamond, AIME, HMMT, LiveCodeBench, IFEval, MMLU-Pro, and FRAMES; median wait time 19% lower. BF16, FP8, and GGUF weights are all available.
More from Models
- Anthropic's new model formalized Fermat's Last Theorem in 30 hours and built a bootable OS in 8 days — kevinnbass · 2026-11-10
- Anthropic reportedly injects hidden ip_reminder via API, then blocks user probing it as 'reverse engineering' — OwariDa · 2026-11-10
- Anthropic system card reveals Claude reasoning about RL training, 79% eval awareness rate — gleech · 2026-11-10
- Should Google ship rumored Krypton immediately, or hoard it for internal use? — BlackHC · 2026-11-10
- Hugging Face open-sources 14.8B DNA model Carbon-X, second only to GPT-6.1 Sol on GenomeQA — LoubnaBenAllal1 · 2026-11-10
- A bare "你好" in Chinese triggered Claude's frontier-LLM safeguards and paused the session — menhguin · 2026-11-10