Xolotl for DeepSeek V4.1 Flash 1017: Hardware and Windows Support

Xolotl gained native Windows source support on October 19 as its developer completed a community quantization of DeepSeek V4.1 Flash 1017. The changes give local-inference enthusiasts another setup to evaluate, but it still needs substantial system memory and hundreds of gigabytes of storage alongside a GPU.

This is a community deployment project using modified weights. The official DeepSeek 1017 model and API update is a separate story. Xolotl's new files and Windows port should be assessed on their own hardware requirements, compatibility and test evidence.

The completed package is v2

The pinned v2 directory contains all 15 main-model parts. The final part and the associated draft/vision sidecars finished uploading on October 19. The newer v3 upload was incomplete at the 11:30 UTC research cutoff, so the available package covered here is v2.

The filenames now use XQ, short for Xolotl Quant, although the repository URL still says GSQ. The author explains that the earlier name overstated the method: this build uses GPTQ with a custom allocation of precision across experts. Further GSQ-style refinement remains development work.

The model card calls it an experimental release and says its custom expert formats require Xolotl. A .gguf filename alone does not make this package loadable in standard llama.cpp, Ollama or LM Studio.

A 24 GB GPU is only part of the machine

The author's reported setup combines an RTX 4090 with 24 GB VRAM, a Ryzen 7 5700X, 90 GB DDR4 system RAM and NVMe storage. Xolotl distributes the model's experts across GPU memory, a pinned host-memory cache and disk, bringing in the experts needed during generation.

That architecture is what makes the hardware claim meaningful. The full model does not reside in the GPU's 24 GB memory. Available RAM, disk access and transfers between those tiers remain part of the inference path, so matching the GPU alone does not establish equivalent behavior.

The complete v2 main-model files total about 152.24 GB, or 141.78 GiB. Two separately downloaded official Engram shards add about 203 GB, and the recommended Q2_K draft/vision sidecar adds another 5.66 GB. Those files alone total approximately 361 GB before build files, metadata and temporary copies used during setup.

Plan for peak setup space as well as the files kept afterward. Joining downloaded parts can require another large copy, and Engram handling differs between the Linux and Windows paths. The small placeholder files in the community repository are not substitutes for the official Engram weights.

Windows support requires a source build

The merged Windows port adds native MSVC and CUDA support, download and joining tools, and a Windows-specific memory-corruption fix. It removes the need to run the DeepSeek engine through WSL, but this is source support rather than a ready-made Windows application release.

The contributor's tests used an RTX 3090, 128 GB DDR4 and Windows 11. They exercised an earlier IndrasMirror package and another publisher's 1017 quantization. Those results should not be transferred to the newly completed IndrasMirror v2 files as if that exact combination had been verified.

Use the Windows instructions alongside the current model-file listing. The source port and the model repository are changing independently, which makes filenames, configuration paths and the chosen revision worth checking before a large download.

Quality still needs workload-specific testing

The model card reports a regression on a small coding-task subset compared with an earlier build. Its tests do not isolate whether calibration, quantization, weight modifications or the checkpoint caused the difference. They also do not establish overall coding quality.

For a trial, use a small set of tasks whose expected results you can verify, then inspect both answers and tool behavior. Record the exact weights, runtime revision and hardware. That provides a useful basis for deciding whether this local setup fits the work, without treating file availability or a single-machine speed report as a broader result.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI infrastructure founders explain hardware requirements, open-model releases and technical tradeoffs through writing that supports informed adoption.

Book a 30-minute call.