EschaLabs has released Escha-W2, a 2-bit quantized build of Qwen3.6-35B-A3B that fits on a single consumer graphics card. Quantization compresses a model’s weights from their original higher precision down to about two bits each, trading some accuracy for a much smaller file. The result is a 35-billion-parameter Mixture-of-Experts model, meaning it routes each token through a handful of specialized subnetworks rather than the whole network, that occupies just 12.3 GB on disk.

The package runs on one 24 GB GPU such as an RTX 4090 or RTX 3090, and EschaLabs says a 16 GB card such as the RTX 5060 Ti also works, with a single tradeoff: less concurrency or less context length, not both at once. It ships as a complete serving stack, wrapping the weights behind an OpenAI-compatible HTTP interface so existing client tools can point at it without modification.

EschaLabs reports 225 tokens per second for a single user on an RTX 4090, and about 1,321 tokens per second of combined throughput at a batch size of 32. A 24 GB RTX 3090 manages 154 tokens per second solo and 390 tokens per second batched, while a 16 GB RTX 5080 hits 212 tokens per second solo. By the company’s own numbers, the 16 GB RTX 5080 outran the 24 GB RTX 3090 by 38 percent, evidence that memory bandwidth matters more than raw VRAM capacity for this workload.

The real question with any 2-bit build is what got lost. EschaLabs compared Escha-W2 against an FP8 version of the same base model (FP8 needs roughly 35 GB, itself a heavier reference point) across six capability tests and reports an average retention of 100.2 percent. Apparent gains on math reasoning and graduate-level science appear in the results table, but EschaLabs’ own footnotes attribute those to evaluation-protocol quirks and run-to-run noise rather than real improvement. The one capacity-bound loss is long-horizon code generation, where LiveCodeBench scores fell to 93.4 percent of the FP8 baseline. Every figure here comes from EschaLabs’ own test harness and its own choice of FP8 as the reference point. No independent lab has verified the comparison.

What this changes is who can run a 35-billion-parameter MoE model without renting cloud GPU time. A developer with one 16 or 24 GB card, the kind found in a high-end gaming rig, can now serve a model of this class locally and privately, at speeds EschaLabs says far outpace a human reader. It does not replace a hosted frontier model for coding-heavy work, where the card’s own numbers show the sharpest degradation, and it does not remove the burden of tuning the serving stack: the model card lists several settings, memory fractions, CUDA graph batch sizes, prefix caching, that silently underperform or crash under careless defaults.

Teams evaluating on-device inference for cost or privacy reasons should treat Escha-W2 as a credible option for general reasoning and tool-calling tasks, but should test their own coding workloads before swapping out a hosted frontier model.

According to EschaLabs’ Hugging Face model card for Escha-W2, with benchmark and hardware figures the company says were measured on 2026-07-27.