# Escha Labs

> Escha Labs builds extreme low-bit compression for large language models. Our first public release, Qwen3.6-35B-A3B-Escha-W2, is a 2-bit build of a 35-billion-parameter Mixture-of-Experts model: 12.3 GB on disk, running on one consumer GPU, decoding 225 tokens/s on an RTX 4090, and holding 100.2% of its FP8 evaluation baseline's unweighted mean across six capability benchmarks. This file indexes the first-party sources; /llms-full.txt carries the complete dataset with measurement conditions attached.

Last updated 2026-07-28. Escha Labs, Inc. — "The LLM Compression Company." Contact ping@eschalabs.com. All weights and runtimes are Apache-2.0.

The headline is not the size. A 35B model at 2 bits has been buildable for years; what usually follows is a large quality loss. The measured result here is that this build stays within run-to-run noise of its FP8 baseline on five of six capability axes, with one real regression in long-horizon code generation (LiveCodeBench v6, 93.4% retention). Size, speed and quality together are the claim: 12.3 GB, 225 tok/s on a single RTX 4090, 100.2% of the baseline's six-axis mean.

Two baselines are in play and they serve different roles. The model is quantized from the BF16 checkpoint, whose safetensors total 71.9 GB in the published Qwen repository — a 5.85x size reduction to 12.30 GB. Quality is measured against an FP8 checkpoint (35.0 GB, a 2.85x reduction) because BF16 does not fit on a 24 GB card at all; FP8 is a full-precision proxy that tracks BF16 on these tasks. FP8 is not itself full precision, and if it sits marginally below BF16 then retention against BF16 would be slightly below the figures quoted here.

Hardware and platform reality, stated up front: NVIDIA GPU with compute capability 8.0 to 12.0 (Ampere through Blackwell), 16 GB VRAM minimum with a reduced context window and 24 GB recommended, Linux x86-64 with glibc 2.28 or newer. There is no macOS build and no Windows build, and WSL2 is untested. All five measured GPUs are desktop cards; no laptop GPU has been measured.

## Primary sources

- [Qwen3.6-35B-A3B-Escha-W2 model card](https://huggingface.co/EschaLabs/Qwen3.6-35B-A3B-Escha-W2): the authoritative document. Full quality tables against the FP8 baseline, the evaluation protocol with per-benchmark thinking mode and token budgets, per-GPU serving figures, the NVIDIA ISL/OSL throughput grid, and an MLPerf-style latency SLA table.
- [escha-runtime-qwen3moe](https://huggingface.co/EschaLabs/escha-runtime-qwen3moe): both serving engines. SGLang versus ZML comparison, long-context and multi-turn measurements, shared requirements, and known limitations.
- [SGLang install and tuning cookbook](https://huggingface.co/EschaLabs/escha-runtime-qwen3moe/blob/main/sglang/INSTALL.md): per-architecture and per-VRAM launch recipes for 16, 24 and 40 GB cards, plus the bounded-thinking budget.
- [Escha Labs](https://www.eschalabs.com/): company site — models, benchmark tables, the compression stack, and contact.

## Complete data

- [llms-full.txt](https://www.eschalabs.com/llms-full.txt): every published number with its measurement conditions — quality per axis, speed on five GPUs, serving under Poisson load, hardware requirements, the install sequence, the four settings that fail silently, and the limitations we publish.

## Optional

- [EschaLabs on Hugging Face](https://huggingface.co/EschaLabs): all published weights and runtimes.
- [Base model Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B): the upstream checkpoint this build quantizes.
- [Contact](https://www.eschalabs.com/#contact): collaboration on model compression, inference acceleration, and research.