The source describes the base as a 763.2B-parameter, 384-expert mixture-of-experts model, with six routed experts active per token and a 1M-token context window.
The releases are IQ1_M at 1.75 bits per weight and 231.32 GB total, IQ2_XXS at 2.00 bits and 248.84 GB, and IQ2_XS at 2.25 bits and 265.83 GB. A shared 110.60 GB Engram lookup-table file is required alongside four backbone shards. The repository recommends IQ2_XXS; its reported task average is 93.54, matching the cited FP8/FP4 base on AIME25, GPQA-Diamond, and LiveCodeBench v6 at reasoning effort 50.
The card says the files run with Strata on one NVIDIA GPU with 12 GB of VRAM by using system memory and memory-mapping the Engram tables from SSD. It also says llama.cpp did not support DeepSeek-V4.1 as of October 23, 2026. The source does not establish independent benchmark replication, real-world throughput, total host-memory requirements, or performance on hardware other than the stated setup.