DeepSeek V4.1 Flash on two DGX Spark

DeepSeek V4.1 Flash on two DGX Spark

Runs DeepSeek V4.1 Flash on two NVIDIA DGX Spark computers at about twice the decode and prefill speed of a standard vLLM setup. A recipe card with no weights, for people serving agents on that hardware.

Made by Elevatormusic on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository is a recipe card for running DeepSeek V4.1 Flash on two NVIDIA DGX Spark computers. It holds no weights. The weights sit in two other repositories, and the full recipe lives on GitHub. The author reports about twice the speed of a standard vLLM setup on the same pair of machines.

Who it suits: It suits people who own two DGX Spark computers and want a local model that handles agent tool calls and extremely long prompts.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Developers building agents | Run tool calls and structured replies against a local model | Forced tool calls and schema-shaped answers
Research | Teams working with long documents | Send prompts up to about 1M tokens | A needle test over a 1,035,020-token prompt
IT | People who host models on their own hardware | Replace a standard vLLM kit | Higher decode speed on the same two computers

If you do not have two DGX Spark computers, you can skip it.

Where to use it:

• Two NVIDIA DGX Spark computers: the pair the numbers were measured on
• The author's TensorFold fork: the engine, on the branch deepseek-v41-tp2
• Agent setups that need forced tool calls or structured output
• Windows, macOS or Linux with Python: only to download the card's files

Title: Jobs it does
Use case | Who | What they get back
Agent tool calls | Agent builders | Tool calls the model is forced to make, and replies shaped to a given format
Questions about long documents | Analysts | A second question on the same document starts in about 1 s instead of 13 s
Several users at once | Small teams | Four requests can run side by side on the pair

How to set it up:

• Two NVIDIA DGX Spark computers
• Python, for downloading the card's files
• The recipe on GitHub, which fetches the two sets of weights at exact versions

1. Install the Hugging Face Hub tools.
2. Download this card's files.
3. Follow the GitHub recipe. It downloads the model weights, 43 files at 2.9 bpw in EXL3, plus two Engram table files, numbers 47 and 48.
4. Use the Engram tables of the 1017 release, never those of the older V4.1 Flash release.

pip install -U huggingface_hub
hf download Elevatormusic/DeepSeek-V4.1-Flash-1017-2x-DGX-Spark-TensorFold-2x-Faster

How to use it: Once the recipe has started the engine on both computers, you send requests to it as you would to any model server. The card names two features that agents lean on.

Title: Features the card names
Tool | What it does
tool_choice | Forces a tool call, either any tool (required) or one named tool
Structured output | Limits the reply to a JSON schema, a regex or a fixed set of options
Prefix cache | Reuses a long prompt, so a repeat question starts faster, even after a restart

hf download Elevatormusic/DeepSeek-V4.1-Flash-1017-2x-DGX-Spark-TensorFold-2x-Faster

Example: You paste a long report and ask "List the open risks in this report." Then you ask a second question about the same report. The card says the second answer starts in about 1 s, where the first long prompt waited about 13 s.

For newcomers and experts: If you are new, think of this as a set of instructions for getting a huge model to run quickly on two NVIDIA computers. The first small thing to try is the speed table on the model page, then the download command above.

If you know the field, the weights are EXL3 at 2.9 bpw and the engine is a TensorFold fork. The card gives prefill of 2,165 tok/s at 32K tokens against 925 for the standard setup, and decode of 56.8 against 27.0 tok/s on prose. The trade-off is that the vLLM prefill figure may include a prefix-cache hit, which would favor vLLM.

Fetch the weights at exact versions
Start the TensorFold engine on two computers
Send prompts, tool calls or schemas
The prefix cache speeds repeat questions

Good to know:

Title: Decode speed with one request
Standard setup, prose: 27.0 tok/s
This recipe, prose: 56.8 tok/s
Standard setup, code: 39.9 tok/s
This recipe, code: 92.2 tok/s

• Scope: the speed results cover the base recipe only, not the optional experimental BF16 overrides the card describes, which come from an uncensored source.
• Context: the longest needle test used 1,035,020 tokens and succeeded, with 13.6 minutes before the first token.
• Concurrency: up to 4 requests at once, with aggregate decode reaching 140 tok/s on code.
• No weights are stored in this repository.
• License: apache-2.0.
• Cost to run: no API key or paid service is named, but you need two DGX Spark computers.
• Data it reaches: it downloads weights from two Hugging Face repositories, and the recipe comes from GitHub.
• Not for: older V4.1 Flash Engram tables, which the card says to avoid.

On Hugging Face, 15 November 2026

• Task: text generation
• Downloads in the last 30 days: 0
• Likes: 9
• License: apache-2.0
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Tags: recipe, tensorfold, exl3, exllamav3, deepseek, deepseek_v41, dgx-spark, gb10, tensor-parallel, tool-calling, structured-output, text-generation

Links

• Model: https://huggingface.co/Elevatormusic/DeepSeek-V4.1-Flash-1017-2x-DGX-Spark-TensorFold-2x-Faster
• Files: https://huggingface.co/Elevatormusic/DeepSeek-V4.1-Flash-1017-2x-DGX-Spark-TensorFold-2x-Faster/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Elevatormusic/DeepSeek-V4.1-Flash-1017-2x-DGX-Spark-TensorFold-2x-Faster

Then follow the usage in its model card. Nothing in this kit runs by itself.