Kvark-Sparkle-1.0-27B

Kvark-Sparkle-1.0-27B

Reasons over images and text with fewer thinking tokens than its base Qwen3.8-27B, so answers arrive in about half the time at similar accuracy, per its card. FP8 and GGUF copies exist.

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: KVARK Sparkle 1.0 is a 27B model from Egzakta that takes images and text and answers in text. It is built on Qwen3.8-27B with the architecture left unchanged. The card reports that it reaches about the same accuracy as its base in roughly half the answer time, because it spends fewer tokens on reasoning.

Who it suits: It suits people who run a large reasoning model and want shorter waits for each answer.

Industry | Who uses it | How they use it | Example
Software and AI teams | Engineers who serve large models | Run the FP8 copy under vLLM | Ask a hard question with thinking on and get a shorter reasoning trace
Research | Researchers studying efficient reasoning | Compare reasoning tokens against Qwen3.8-27B | Check the GPQA Diamond, AIME and LiveCodeBench figures
Local AI | Enthusiasts with large GPUs | Use a GGUF copy with llama.cpp | Pick a quant between Q8_0 and IQ4_XS

If you need a small model for a laptop, you can skip it, since even the smallest GGUF copy is 15.3 GB.

Where to use it:

• Python with the transformers library, for the BF16 weights in this repository
• vLLM, using the separate FP8 copy
• llama.cpp, using the separate GGUF copies
• Windows, macOS or Linux with Python, to download the files

Weights | Runs in | Size
BF16 with an MTP head | transformers | 52 GB
FP8 with an MTP head | vLLM | 29 GB
GGUF, Q8_0 to IQ4_XS | llama.cpp | 15.3 to 29.0 GB

How to set it up: What you need first:

• Python and the Hugging Face command line
• Enough GPU memory for the 54.7 GB of BF16 weights, or a smaller FP8 or GGUF copy
• About 13.7 GB if you run a 4-bit version, which is an estimate

1. Install the Hugging Face tools with pip.
2. Download the model files into your cache or a folder.
3. Load them with transformers, or fetch the FP8 or GGUF copy for vLLM or llama.cpp.

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

How to use it: Send the model an image together with a question, and leave thinking on, which is how the card's benchmarks ran. The runs used temperature 1.0, with top_p at 0.95 and top_k at 20.

Tool | What it does
transformers | Loads the BF16 weights from this repository
vLLM | Serves the FP8 copy
llama.cpp | Runs the GGUF copies

Example: You give it a chart image and ask which quarter grew fastest, with thinking on. The card's numbers suggest a shorter reasoning trace than Qwen3.8-27B would use, with a similar answer quality.

For newcomers and experts: A good first step is to download the files, load the model with transformers and send one image with a short question. Start with the FP8 or GGUF copy if your GPU is small.

For experts, the model keeps the Qwen3.8-27B design: 64 layers, of which 48 use linear attention and 16 use full attention. Full attention has 24 query heads, 4 key-value heads and a head dimension of 256. The hidden size is 5,120 and the vocabulary holds 248,320 tokens. Only the 16 full-attention layers store a KV cache, which takes 64 KiB per token in BF16 and 32 KiB in FP8. One MTP layer from the base model supports speculative decoding, and the vision encoder has 27 layers.

An image and a question go in
The model reasons with thinking on
It spends fewer reasoning tokens than Qwen3.8-27B
The final answer arrives sooner

Good to know:

Title: Median time to answer
Qwen3.8-27B: 28.1 s
KVARK Sparkle 1.0: 13.7 s

Title: Median reasoning tokens
Benchmark | Sparkle | Qwen
GPQA Diamond | 3,122 | 7,854
AIME | 5,502 | 14,318
LiveCodeBench | 6,664 | 19,950
IFEval | 470 | 714
MMLU-Pro | 429 | 727

• Evaluation: each model used FP8 weights with thinking enabled and a 64k token limit, and the differences are paired over items.
• Context: 262,144 tokens natively, but the card evaluated only up to 64k.
• Language: English is the one language listed.
• License: apache-2.0
• Cost to run: no API key or paid service named. You need a machine with enough GPU memory.
• Data it reaches: it downloads files from Hugging Face and then runs on your own machine.
• Not for: none named.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 315
• Likes: 15
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 8
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, reasoning, efficient-reasoning, qwen3.8, kvark, conversational, model-index

Links

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

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