miaai-lab/deepseek-v4.1-flash-1017-2x-dgx-sparks is a deployment recipe that serves DeepSeek-V4.1-Flash-1017 EXL3 based on TensorFold on two DGX Sparks. The repository documents model checkpoints and configuration values, plus preparation and run scripts. No separate latest release version is recorded in FACTS.
The README describes an OpenAI-compatible API, image input, tool calling, reasoning mode switching, and DSpark speculative decoding. Speculative decoding is a method that generates candidate tokens first and then verifies them. The context window covers settings up to 1,048,576 tokens, but the actual measurements include the 524,288-token setting. For the full-window setting, it records running a 100,000-token input. Configuration checks, build results, and performance measurements under different tools and conditions are also presented separately.
However, it states that image input and tool calling were not verified on that pair of devices. Some compute plans also were not re-run on the actual 39-shard model. The performance figures are the project's own measurements, and no evidence of independent reproduction is presented. So rather than accepting all the features and settings listed in the document as verified performance, one should look at the measurement conditions and the scope of verification together.