Unsloth ships DeepSeek-V4.1-Flash-1017 dynamic GGUFs: a 250 GB 1-bit quant runs on 128 GB+ RAM, with the Engram table offloaded to SSD

DeepSeek-V4.1-Flash-1017, released by DeepSeek on Oct 17 under the MIT license (DeepSWE 77.0, vendor-reported), is one of the most watched open-weight models, but its lossless weights exceed 500 GB, which keeps most developers on the API. On Oct 19 Unsloth published dynamic GGUF quants and a run-locally guide. According to Unsloth, the model combines a 552B backbone with 196B Engram parameters. Engram is a conditional memory where n-grams of up to four tokens touch only a few table rows per token, so the 110.6–202.8 GB table can stay on SSD via mmap and load rows on demand, cutting RAM/VRAM needs but not disk. Unsloth-reported agreement with DeepSeek's reference weights: UD-IQ1_M (250.4 GB) 88.2% top-1 on long17k and 87.1% on wiki1100; UD-IQ1_S (233.0 GB) 85.1%/82.8%; UD-IQ3_XXS (394.7 GB) 95.1%/94.6%; UD-Q4_K_XL (500.7 GB) 98.5%/98.2%. Top-1 agreement is not task accuracy, and no quantized task benchmarks were published. Unsloth also claims ~3x faster inference than other CPU/VRAM/ngram-focused repos. To try it, budget ~128–145 GB of RAM+VRAM for 1-bit or ~256–290 GB for 3-bit, download unsloth/DeepSeek-V4.1-Flash-1017-GGUF via Unsloth Desktop or build Unsloth's llama.cpp PR (mainline is not yet supported), use temperature 1.0 and top_p 1.0 (reasoning_effort defaults to 75), and serve it with unsloth run to connect Claude Code or Codex locally.