DeepSeek-V4.1-Flash-1017-Basalt

DeepSeek-V4.1-Flash-1017-Basalt

Packs DeepSeek V4.1 Flash into single model files for the Basalt engine, which runs it on one NVIDIA Blackwell GPU with system RAM covering the experts that do not fit. Supports images and up to 1,048,576 tokens of context.

Made by jesdga 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: Basalt runs language models locally on NVIDIA Blackwell cards and serves them through APIs that mimic those of OpenAI and Anthropic. This repository supplies model files for it. Each file bundles one Unsloth quantization of DeepSeek-V4.1-Flash-1017 with the weights, Engram table, vision projector, tokenizer and DSpark drafter. You need Basalt 0.3.0 or later.

Who it suits: It suits owners of recent NVIDIA workstation cards who want to run a huge open model locally behind a familiar API.

Industry | Who uses it | How they use it | Example
Local AI hobby | Enthusiasts | Run a large DeepSeek model on a home workstation | Chatting with it on one Blackwell card
Software development | Developers | Point OpenAI-style or Anthropic-style clients at a local server | Running long agentic coding sessions
Vision tasks | App builders | Send images to the model through the engine | Asking about a screenshot with the vision flag on
Quantization research | Researchers | Compare each pack against a BF16 reference | Reading the perplexity and KLD tables

You can skip it if you have no NVIDIA Blackwell GPU, since the engine targets that architecture.

Where to use it:

• The Basalt engine on NVIDIA Blackwell GPUs
• A single GPU, with an optional second GPU that holds extra experts
• Spare system memory, which takes the experts that exceed what the GPU can hold
• Windows, macOS or Linux with Python, only to download the files

Use case | Who | What they get back
Local chat and coding | Developers | Text answers through an OpenAI-style or Anthropic-style API
Image questions | App builders | Answers about images encoded inside the engine
Long context | Analysts | Contexts up to 1,048,576 tokens

How to set it up: You need a Blackwell GPU, the Basalt engine built from its own README, and disk space for the pack, 260 GB or 285 GB.

• An NVIDIA Blackwell GPU, such as the RTX 5090 used in the card's tests
• Basalt 0.3.0 or later, built following its README
• 260 GB for the smaller pack or 285 GB for the larger one
• Python and huggingface_hub to download a pack

1. Build the Basalt engine by following its README.
2. Download one pack file into a models folder.
3. Start the server on that file.

pip install -U huggingface_hub

hf download jesdga/DeepSeek-V4.1-Flash-1017-Basalt DeepSeek-V4.1-Flash-1017-UD-IQ1_M.basalt --local-dir models   # or the UD-IQ2_XXS file

How to use it: You serve a pack with basalt serve. Add flags for a second card, a longer context or image input. You do not set the prompt format yourself, because the file carries it, along with how tool calls look and how thinking is toggled.

bin/basalt serve models/DeepSeek-V4.1-Flash-1017-UD-IQ1_M.basalt --vision dynamic

Tool | What it does
--vision dynamic | Turns on image input, encoded inside the engine
--max-context | Sets a longer context, up to 1,048,576 tokens
--gpu 0,1 | Uses a second card
--spec 0 | Turns off the DSpark drafter

Example: You start the server with the smaller pack and the vision flag, then send a picture and a question from an OpenAI-style client. The engine encodes the image itself and the model answers in text.

For newcomers and experts: If you are new to this, treat it as a giant model squeezed into one file for a single high-end graphics card. A first small try is to download the smaller pack and start the server with the command above.

For people who know the field, two packs trade speed for quality. UD-IQ1_M is faster and keeps more of the model in RAM, while UD-IQ2_XXS sits closer to BF16 but slows down when experts spill to disk. DSpark drafts three tokens at a time, and the main model verifies them, so output stays the model's own.

Title: Prose speed in tokens/s
UD-IQ1_M: 94 tokens/s
UD-IQ2_XXS: 45 tokens/s

You download one pack file
Basalt loads it on the Blackwell GPU
Experts that do not fit stay in system RAM
The API answers text, or images with vision on

Good to know: The card reports speeds and quality from its own runs, and says the numbers move with each release.

• License: MIT, covering the DeepSeek model, Unsloth's quantization and the Basalt engine
• Cost to run: none named, though you need a Blackwell GPU and large disk and RAM
• Data it reaches: local files and system RAM, with experts that spill to disk
• Not for: GPUs other than NVIDIA Blackwell, or Basalt versions before 0.3.0

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 11
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Tags: basalt

Links

• Model: https://huggingface.co/jesdga/DeepSeek-V4.1-Flash-1017-Basalt
• Files: https://huggingface.co/jesdga/DeepSeek-V4.1-Flash-1017-Basalt/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download jesdga/DeepSeek-V4.1-Flash-1017-Basalt

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