KVARK Sparkle 1.0 (27B) FP8

KVARK Sparkle 1.0 (27B) FP8

Runs a 27B model that reads images and text, in FP8 for vLLM. It reasons with far fewer tokens than a Qwen model of the same size. Made for people who host their own models.

Made by Egzakta 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 is the FP8 copy of KVARK Sparkle 1.0, a 27.8B model from Egzakta that reads images and text and writes text. It is packaged for vLLM and keeps the multi-token prediction head. Egzakta reports that it reasons with far fewer tokens than a 27B Qwen model while staying close on accuracy.

Who it suits: It suits people who host their own language models and want shorter thinking time on hard questions. The card measures science, math, coding and instruction tasks.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Engineers who host models | Serve it with vLLM on one big GPU | A coding helper that answers hard problems
Research | Machine learning researchers | Compare reasoning length against another model | Check token use on science questions
Education | Teams building tutoring tools | Run step by step reasoning on math problems | A math practice helper

If you only want a small model on a laptop, you can skip it, since the files take 29 GB.

Where to use it:

• vLLM: the serving route the card describes, with the prediction head on
• A single H200 GPU: it serves these weights with about 128k context
• Windows, macOS or Linux with Python: for downloading the files
• Transformers: the library tag on the model page

Title: Jobs it does
Use case | Who | What they get back
Science questions | Researchers | Written reasoning, then a final answer
Math problems | Students and teachers | A step by step solution
Code tasks | Developers | Code written with shorter reasoning than the Qwen model

How to set it up:

• Python on Windows, macOS or Linux
• A GPU with room for 29 GB of weights, such as one H200
• vLLM to serve the model, the card measured speed on version 0.30

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Follow the serving steps in the model card to start vLLM.

pip install -U huggingface_hub
hf download Egzakta/Kvark-Sparkle-1.0-27B-FP8

How to use it: Start vLLM with the serving steps in the model card, then send prompts with thinking turned on, as the card's tests did. The download command fetches the files first.

hf download Egzakta/Kvark-Sparkle-1.0-27B-FP8

Example: You send a hard question, such as "What is the chance of drawing two aces from a deck?" The model writes out its reasoning first and then gives the answer. On AIME problems the card reports a 62% cut in median reasoning tokens against the Qwen model.

For newcomers and experts: If you are new, think of this as a large model that thinks before it answers, stored in 8-bit floating point so the files stay near 29 GB. The first small thing to try is the two download commands, then open the model card for its serving steps.

If you know the field, the weights are F8_E4M3 and the prediction head lets the server draft several tokens per step. The card gives a mean acceptance length of 2.53 against 2.50 for the Qwen model. The trade-off is accuracy. Sparkle scored 82.7 against 77.3 on LiveCodeBench and slightly lower on the other four tests, and the 95% intervals touch or cross zero on all five.

Download the FP8 files
Serve them with vLLM
Send a prompt with thinking on
The model reasons, then answers
The prediction head speeds decoding

Good to know:

Title: Median seconds per answer on one H200
Sparkle with MTP: 13.7 s
Qwen with MTP: 28.1 s
Qwen without MTP: 49.1 s

• Speed test: 20 fixed questions, one request at a time, so it says little about many users at once.
• Language: the record lists English only.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named, but you need a GPU such as one H200.
• Data it reaches: you download the files from Hugging Face, and the card names nothing else.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 27.8 GB in F8_E4M3; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 24,270
• Likes: 5
• License: apache-2.0
• Languages: en
• Based on: Egzakta/Kvark-Sparkle-1.0-27B
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, fp8, vllm, reasoning, kvark, conversational

Links

• Model: https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-FP8
• Files: https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-FP8/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Egzakta/Kvark-Sparkle-1.0-27B-FP8

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