Escha-W2
by Escha Labs · Model · Language & code models
2-bit build of Qwen3.8-27B shrinks the 50 GB model to about 10 GB, with a fine-tuning pass to recover quality
AccessOpen weightsNeeds~24 GBDownload10.2 GBSize27B parametersLicenseApache-2.0
Evidence
- Aug 21Claimed to tie or beat the FP8 base model on major benchmarks: Commonsense-6 79.25 vs 77.96, GPQA-Diamond 88.38 vs 88.89, LiveCodeBench v6 86.81 vs 85.16
- Aug 21Served with Escha's own SGLang build; used 36 GB VRAM with a large KV cache, reducible to under 24 GB