DeepSeek-V4.1 TensorFold Spark
A recipe for serving DeepSeek-V4.1-Flash-1017 on two NVIDIA DGX Sparks with the TensorFold engine behind an OpenAI-compatible API.
At a Glance
Apache-2.0 licensed recipe, code, patches, scripts, benchmarks and docs for serving DeepSeek-V4.1-Flash on 2x NVIDIA DGX Spark. TensorFold, adapted kernels, the base image and model weights keep their own separate terms.
Engagement
Available On
Listed Nov 2026
About DeepSeek-V4.1 TensorFold Spark
This repository by jayleaton provides a recipe for serving DeepSeek-V4.1-Flash-1017 on two NVIDIA DGX Spark machines behind an OpenAI-compatible API. It ships an EXL3 2.8 bpw model pack (flash17) and a pinned upstream TensorFold native build. The project is marked as a work in progress and measured on one pair of Sparks.
What It Is
The project is a self-hosted LLM serving kit. It combines the DeepSeek-V4.1-Flash-1017 model, a 203.8 GB EXL3 quantized pack, and the TensorFold engine running with tensor parallelism across two DGX Sparks (GB10, 128 GB each) linked over a 200 Gb/s CX7 connection. The serving profile covers 1M context, four slots, native image input, tool calling, structured output and a built-in Python tool.
Setup Path
The quick start builds the engine image and a runner image with scripts, downloads the model pack from Hugging Face on both nodes, and configures per-node paths from an example env file. Engram shards must be packed from the 1017 source, and the recipe requires a 2.1M-token pool. The README suggests a 2100-2200 MHz GPU clock cap for temperature control and a sanity check (17 * 23 = 391) before serving traffic. A rollback path to the previous q28-v2 pack is documented.
Measured Results
The README reports results on 356 matched items across suites such as HumanEval+, MBPP+, MMLU-Pro STEM, MATH-500, LiveCodeBench and BFCL, compared with another 2.9 bpw pack and the earlier q28-v2 pack. It notes the samples are small and show no statistically significant difference. It reports single-stream short decode of 87.9 tok/s at the recommended profile. Full engine, pack and device qualification for the current pin is stated as still pending.
Current Status
The repository labels itself a work in progress and points to a known-issues document. Code, patches, scripts and docs are under Apache License 2.0, while TensorFold, the base image and the model weights keep their own terms.
Community Discussions
Be the first to start a conversation about DeepSeek-V4.1 TensorFold Spark
Share your experience with DeepSeek-V4.1 TensorFold Spark, ask questions, or help others learn from your insights.
Pricing
Open Source
Apache-2.0 licensed recipe, code, patches, scripts, benchmarks and docs for serving DeepSeek-V4.1-Flash on 2x NVIDIA DGX Spark. TensorFold, adapted kernels, the base image and model weights keep their own separate terms.
- Apache License 2.0 for project code, patches, scripts, benchmarks and docs
- Self-hosted on 2x NVIDIA DGX Spark (GB10, 128 GB each)
- OpenAI-compatible API
- Vision, tool calling, structured output, 1M context
- TensorFold, adapted kernels, base image and model weights keep their own terms (docs/LICENSING.md)
Capabilities
Key Features
- Serves DeepSeek-V4.1-Flash-1017 on two NVIDIA DGX Sparks with tensor parallelism
- OpenAI-compatible API
- EXL3 2.8 bpw quantized pack (flash17)
- 1M context with four slots
- Native image input (vision)
- Tool calling and structured output
- Built-in Python tool via compute lane
- Pinned upstream TensorFold native build
- Rollback profile to previous q28-v2 pack
- Benchmark and preflight HTTP tooling
