---
type: "post"
slug: "eschalabs-releases-escha-w2-model-on-hugging-face"
title: "Escha-W2 compresses a 35B MoE into a local serving footprint"
description: "The Apache-2.0 two-bit Qwen3.6 derivative is roughly 12.3 GB and targets machines with 16–24 GB of memory."
retrieval_nugget: "The Apache-2.0 two-bit Qwen3.6 derivative is roughly 12.3 GB and targets machines with 16–24 GB of memory."
published_at: "2026-08-23"
updated_at: "2026-08-23"
record_date: "2026-08-23"
date_kind: "scheduled_at"
topics: ["local-models","quantization","open-models","moe"]
entities: ["huggingface.co"]
editorial_format: "brief"
basket_id: "64af3bcb-1c2d-42a9-a664-91510a61d75a"
basket_revision: 1
source_urls: ["https://huggingface.co/EschaLabs/Qwen3.6-35B-A3B-Escha-W2"]
visual_decision: "text_only"
recovery_incident: "NR-2026-08-15-HERMES-SITE-COPY"
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:eschalabs-releases-escha-w2-model-on-hugging-face"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:b32c28e9854bcfb79593596eb60e7ac0562f45a1fd166e08980b6a7a35439034","reviewed_at":"2026-08-15T20:30:00.000Z","source_evidence_count":1,"verified_claim_count":2}
routes: {"html":"https://newruntime.com/posts/eschalabs-releases-escha-w2-model-on-hugging-face/","markdown":"https://newruntime.com/posts/eschalabs-releases-escha-w2-model-on-hugging-face.md","json":"https://newruntime.com/posts/eschalabs-releases-escha-w2-model-on-hugging-face.json"}
---

# Escha-W2 compresses a 35B MoE into a local serving footprint

## Retrieval answer

The Apache-2.0 two-bit Qwen3.6 derivative is roughly 12.3 GB and targets machines with 16–24 GB of memory.

EschaLabs released Escha-W2, a two-bit quantization of Qwen3.6-35B-A3B under Apache 2.0. The main artifact is roughly 12.3 GB, with 16 GB listed as a minimum memory target and 24 GB recommended. The underlying MoE uses 256 experts.

The model card documents SGLang for tool use, concurrency, structured output, and reasoning, plus a standalone ZML route and an MLX variant for Apple Silicon. EschaLabs reports about 5.6 times compression with mostly retained quality, while identifying coding as a noticeable weakness.

That makes the release a candidate local executor for bounded tasks, not a free substitute for a frontier service. Evaluation should focus on the exact tool schema, quantization regressions, and memory pressure of the intended runtime. The open license and local footprint make that testing feasible.

## Source and serving instructions

- [Hugging Face model card](https://huggingface.co/EschaLabs/Qwen3.6-35B-A3B-Escha-W2)