Qwen3.6 35B A3B Escha

Qwen3.6 35B A3B Escha is EschaLabs' Escha-W2 build of Alibaba's Qwen3.6 35B-A3B, a mixture-of-experts model with 35 billion total parameters, 3 billion active per token, and 256 experts. EschaLabs compresses the expert weights to a mixed 2-bit and 3-bit format per projection and keeps the dense layers in int8, bringing the model to 12.3 GB on disk so it runs on a single consumer GPU with 24 GB of memory. The model card reports the build keeping 100.2% of an FP8 baseline's average score across six benchmark axes and 100.8% on the BFCL tool-use benchmark, with LiveCodeBench at 93.4% the one clear gap, and a KL divergence of 0.0694 nats against the BF16 original with 89.9% top-1 token agreement, so behavior tracks the full model closely. Thinking mode can be toggled per request and function calling is supported; this build is served as a text-only model, without the multimodal input of the base Qwen3.6 35B-A3B. Released under the Apache 2.0 license, it is aimed at running Qwen3.6-class reasoning and agent tasks on modest hardware, and suits resource-constrained agent loops, extraction, and tool-calling workloads where a compact footprint matters more than the last point of accuracy.