Xolotl is a C++ inference engine that extends Strata to run large MoE models on a single GPU. Its previous name was Strata-DSV4.1. It serves DeepSeek-V4.1-Flash 552B, GLM-5.3-Flash, Qwen3.8-Flash-Next, and Qwen3.8-27B through one OpenAI/Anthropic-compatible server. The repository is MIT licensed.
The core is a tiered approach that splits the model's MoE experts across VRAM, a fixed RAM space, and NVMe. The experts needed for each token are moved in advance, which lets it handle models larger than GPU memory. Mixed-precision quantization, speculative decoding, and a 4-bit KV cache for 512K context are applied on top. According to the README, on an RTX 4090 24 GB with 90 GB DDR4, DeepSeek-V4.1-Flash recorded 26.8 tok/s decode, Qwen3.8-Flash-Next 127.8 tok/s, and Qwen3.8-27B 98.5 tok/s. The README also cites HumanEval 21/21, tool call 10/10, and agentic 3/3, and states that the DeepSeek weight size is 152 GiB. It also says that roughly 71,000 lines were added to the Strata-based code.
However, measurements for GLM-5.3-Flash have not been presented yet. The stated operating conditions are Linux and an NVIDIA GPU with 24 GB or more. The repository has 32 stars, and it is hard to judge the actual scope of use or maintenance status from public information alone. The reported performance numbers are also best treated as results recorded in the README.