Qwen3.8-27B fine-tuned to think 58% less at near-identical accuracy
Emerged finds new products worth a look, daily — what’s new today
KVARK Sparkle 1.0 (27B)
A Qwen3.8-27B fine-tune that reaches the same answers in about half the time
Is this yours? Claim itIf reasoning latency or token bills are your bottleneck, this Apache-2.0 drop-in roughly halves thinking time on a single GPU with near-identical scores.
A fine-tune of Qwen3.8-27B from Egzakta's AI lab, trained to stop overthinking. Two LoRA lines, one rewarded for short correct answers and one for breadth, were averaged and briefly tuned again. On one H200 the median answer took 13.7 s against Qwen's 28.1 s, with 34-67% fewer reasoning tokens. Accuracy is close: level on GPQA and MMLU-Pro, ahead on LiveCodeBench, slightly behind elsewhere. The timing run used only 20 questions. BF16, FP8 and GGUF builds, Apache 2.0.
More like this
A 0.6B decision model that plays ViZDoom, mazes and Snake without decoding
An open 9B model, dataset and training recipe for fast typed decisions over text
A tuned vLLM stack that serves Qwen3.8-27B fast on one RTX 3090
Open 0.8B to 9B decision models for choice, yes/no and rating questions
Exact, shared-context inference server for small Jev-style decision models