ProCreations' Grug-27B cuts token use by 90%, beats Qwen 3.6 27B
A 90% token reduction makes a 27B model run like a 30 tps beast on old laptops.
ProCreations has quietly uploaded a new model called Grug-27B to Hugging Face, and early benchmarks suggest it could be a game-changer for running large language models on consumer hardware. Based on the Qwen 3.6 27B architecture, Grug-27B reportedly outperforms the original in standard tests while slashing the number of tokens required for inference by more than 90%. That level of token reduction is unprecedented for a model of this size and points to aggressive pruning or quantization techniques combined with a novel reasoning pipeline. The claim has not yet been independently verified by the community, but the uploaded results have sparked significant interest.
If the benchmarks hold up, the practical impact is enormous. A 27B parameter model typically runs at around 3 tokens per second on an older laptop GPU. With a 90% reduction in tokens needed for the thinking phase, the effective throughput during reasoning jumps to roughly 30 tps — making complex chain-of-thought tasks feel nearly instantaneous. This could finally make locally-run, high-quality reasoning models viable for professionals without expensive hardware. The community is eagerly awaiting replication tests and a detailed technical report from ProCreations.
- Grug-27B outperforms Qwen 3.6 27B original on benchmarks while using 90% fewer tokens
- Effective inference speed increases from 3 tps to ~30 tps for the reasoning phase
- Model uploaded by ProCreations on Hugging Face; not yet independently verified
Why It Matters
If true, Grug-27B makes large reasoning models practical on older laptops, democratizing access to advanced AI.