Egzakta AI Lab releases KVARK Sparkle 1.0, a Qwen3.8-27B-based model with 262,144 tokens of native context
Egzakta AI Lab's first open weight release, KVARK Sparkle 1.0, carries the Qwen3.8 27B architecture unchanged with 262,144 tokens of native context and is lice…
Published on MyPrivateClaw
Nov 3, 2026, 2:47 AM UTC
Coverage date
Nov 2, 2026
Last updated
Nov 3, 2026, 2:47 AM UTC
News summary
KVARK Sparkle 1.0 keeps the Qwen3.8 27B architecture unchanged: 27.8B parameters in BF16, 64 layers of which 48 use linear attention and 16 use full attention, and a native context of 262,144 tokens. The model is licensed under Apache 2.0, inherited from Qwen3.8 27B per the model card, with commercial use permitted and no revenue cap. Weights are published in BF16 (52 GB, with MTP head), FP8 (29 GB, for vLLM), and GGUF from Q8 0 to IQ4 XS (15.3 to 29.0 GB); supported runtimes include vLLM, llama.cpp, Transformers, and SGLang. In the releasing team's self reported evaluation, measured under FP8 weights with thinking enabled, a 64k cap, and single request execution on one H200, KVARK Sparkle 1.0 uses up to 67% fewer reasoning tokens than Qwen3.8 27B across five benchmarks (e.g., 6,664 vs 19,950 on LiveCodeBench v6). Separately, a median timing test of 20 fixed questions using vLLM 0.30 re…