KVARK Sparkle 1.1 27B GGUF

KVARK Sparkle 1.1 27B GGUF

Runs the 27B KVARK Sparkle 1.1 model from GGUF files in several sizes, with optional image input and speculative decoding. For people running models locally with llama.cpp, Ollama or LM Studio.

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 repository holds GGUF files of KVARK Sparkle 1.1, a 27B model from Egzakta, packaged for llama.cpp, Ollama and LM Studio. Every file keeps the model's multi-token prediction head. The quantized files were built with an importance matrix computed on Egzakta's own fine-tuning data, which is included. A separate vision projector file adds image input.

Who it suits: People who run large models on their own hardware and want a choice of file sizes, from full precision down to about 15 GB.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local 27B model through llama.cpp | A private assistant behind an OpenAI-style endpoint
Document work | Analysts | Ask questions about documents without sending them out | A question about a contract on a local machine
Local AI | Hobbyists | Pick the file that fits their graphics card | The Q4_K_M file on a 24 GB card
Research | Quantization testers | Compare file sizes and memory needs across quant types | KV cache memory at several context lengths

You can skip it if you want the original unquantized model for training, since these files are for running it.

Where to use it: It runs in programs that open GGUF files, on Windows, macOS or Linux.

• llama.cpp, with the llama-server program
• Ollama
• LM Studio

Use case | Who | What they get back
Local chat | Developers | Answers from the model, with its thinking returned in a separate field
Image questions | Hobbyists | Replies about a picture, once the vision file is loaded
Long context | Analysts | Room for contexts up to 262,144 tokens, memory permitting

How to set it up: You need a GGUF runtime and enough memory for the file you choose, from 15.3 GB up to 54.7 GB.

• A runtime such as llama.cpp, Ollama or LM Studio
• Python with pip, if you fetch the files with the Hugging Face tool
• Memory for the chosen file plus the KV cache for your context length

1. Pick a file. The card recommends Q4_K_M for 24 GB cards.
2. Download it, along with the vision projector file if you want images.
3. Start llama-server with the command in the next section.

pip install -U huggingface_hub
hf download Egzakta/Kvark-Sparkle-1.1-27B-GGUF

How to use it: The command below pulls the Q4_K_M file, sets a 65,536 token context and turns on speculative decoding with the included prediction head. Afterwards a local server on port 8080 accepts chat requests in the OpenAI format and returns the thinking in its own field.

llama-server -hf Egzakta/Kvark-Sparkle-1.1-27B-GGUF:Q4_K_M -c 65536 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --temp 1.0 --top-p 0.95 --top-k 20

Tool | What it does
-m and --mmproj | Load local model and vision projector files
--no-mmproj | Text only, skipping the vision projector
-ctk q8_0 -ctv q8_0 | Roughly halves the KV cache memory
--reasoning off | Answers without thinking first

Example: Start the server, then send a chat request about a document you pasted in. The reply comes back as the answer, and the model's reasoning arrives in its own field called reasoning_content, so your app can show or hide it.

For newcomers and experts: A GGUF file is a model packed for running on your own computer. Smaller files use less memory but lose a little accuracy. A first small try is to run the command above with the Q4_K_M file and chat with it.

For experts, the card lists six text files from BF16 down to IQ4_XS, plus an F16 vision projector. The KV cache grows with context, from 2 GiB at 32,768 tokens to 16 GiB at 262,144 in f16, and about half that at q8_0. The linear-attention state stays under 0.2 GB at any length. The trade is file size against quality, which only your own tests can settle.

Pick a GGUF file that fits your memory
Start llama-server with the model
Optionally load the vision projector
Send chat requests on port 8080
Read the answer and the reasoning separately

Good to know: The card's benchmark image describes the original model, and it says those numbers were not measured with these GGUF files.

• Language: The record lists English and Serbian.
• Fewer thinking tokens: The image reports lower median reasoning tokens than Qwen3.8-27B, for example 4,471 against 7,670 on GPQA Diamond.
• Memory: Setting the KV cache to q8_0 roughly halves its size.

Title: File size of each text GGUF in GB
BF16: 54.7 GB
Q8_0: 29.0 GB
Q6_K: 22.4 GB
Q5_K_M: 19.5 GB
Q4_K_M: 16.8 GB
IQ4_XS: 15.3 GB

Title: Median reasoning tokens, Sparkle 1.1 against Qwen3.8-27B
GPQA Diamond Sparkle 1.1: 4,471 tokens
GPQA Diamond Qwen3.8-27B: 7,670 tokens
AIME 2026 Sparkle 1.1: 8,732 tokens
AIME 2026 Qwen3.8-27B: 14,822 tokens
HMMT Sparkle 1.1: 13,540 tokens
HMMT Qwen3.8-27B: 21,182 tokens

• License: Apache 2.0.
• Cost to run: none named, the source lists no API key, account or paid service, though the files need a machine with a lot of memory.
• Data it reaches: the prompts and images you send it, plus the files the command downloads from Hugging Face.
• Not for: reading the card's benchmark figures as results of these GGUF files, since the card says they were measured on the original model.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: gguf
• Parameters: 27.3B
• Weights: about 13.7 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 610
• Likes: 3
• License: apache-2.0
• Languages: en, sr
• Based on: Egzakta/Kvark-Sparkle-1.1-27B
• Tags: gguf, llama.cpp, imatrix, reasoning, agentic-rag, kvark, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF
• Files: https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Egzakta/Kvark-Sparkle-1.1-27B-GGUF

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