Delta Inference Engine v0.93 runs DeepSeek-V4.1-Flash on a single RTX 3090

On November 6, 2026, at 02:23, a post appeared on Reddit's LocalLLaMA community announcing that the Delta Inference Engine had been updated to v0.93. According to the post, the release adds support for 64/96/128 GB of RAM and fully supports the new Unsloth quants. With 192 GB of RAM, an Unsloth IQ2 quant still sustains 41 tok/s, while first token latency drops by roughly 40%. The author attributes the speed to a new native reader for the Engram tables and some careful CPU kernel work. Beyond that, v0.93 introduces EXL3 support, basic multi-GPU capability, and public CI. The poster frames the update as a response to requests for more features and more benchmarks, and says that compared with the old default model on their 99-question benchmark it went ... but the excerpt available cuts off there, so the actual benchmark numbers are not in the material. This is the only report received so far, and there is no follow-up information yet. It should also be noted that the post title references running DeepSeek-V4.1-Flash on a single 3090, while the body text describes the v0.93 features listed above without confirming that particular configuration.