The Xolotl inference engine, built on the Strata framework, enables running the full 552‑parameter DeepSeek‑V4.1‑Flash model on a single RTX 4090 by distributing weights across VRAM, pinned system RAM and NVMe storage. It achieves approximately 27 tokens/second decode speed and 180–320 tokens/second prefill with a 512 K token context window. An uncensored 1017‑quant version is expected to be released in a few days.