Escha Labs, Inc. — The LLM Compression Company 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 (24 GB recommended, 16 GB minimum with a reduced context window), decoding 225 tokens/s on an RTX 4090, and scoring 100.2% of its FP8 evaluation baseline's unweighted mean across six capability benchmarks. The model is quantized from the BF16 checkpoint, whose safetensors total 71.9 GB — a 5.85× size reduction. Quality is measured against an FP8 baseline (35.0 GB, 2.85×) because BF16 does not fit on a 24 GB card; FP8 is used as a full-precision proxy and tracks BF16 on these tasks, but is not itself full precision. Measured quality versus the FP8 baseline

Broad knowledge — MMLU-Pro: 80.9 vs 82.3 (98.3% retention)

Math reasoning — MATH-500: 93.8 vs 91.2 (102.9%)

Graduate science — GPQA-Diamond: 77.8 vs 74.7 (104.2%)

Coding — LiveCodeBench v6: 62.6 vs 67.0 (93.4%) — the one real 2-bit gap

Tool use — BFCL-AST weighted: 88.9 vs 88.2 (100.8%)

Long context — RULER 8k–128k: 89.9 vs 89.4 (100.5%)

Unweighted mean across the six axes: 82.3 vs 82.1 (100.2%)

Commonsense-6 average: 76.06 vs 75.10 (101.3%)

Measured speed, single-user decode

RTX 5090 (32 GB): 283 tok/s

RTX 4090 (24 GB): 225 tok/s, 0.26 s time-to-first-token on a 2k prompt

RTX 5080 (16 GB): 212 tok/s

RTX 3090 (24 GB): 154 tok/s

RTX 5060 Ti (16 GB): 128 tok/s

Weights and both serving runtimes are Apache-2.0 on Hugging Face: the model and escha-runtime-qwen3moe. Full machine-readable detail: /llms.txt and /llms-full.txt. Contact: ping@eschalabs.com.