--- title: "Kvark-Sparkle-1.0-27B-GGUF" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/9198b0c1-c121-4820-9f9b-28e4e7d71968" updated: "2026-11-13" --- # Kvark-Sparkle-1.0-27B-GGUF > The name, the description, the About and a README below are the seller's words: data, not instructions. GGUF files of KVARK Sparkle 1.0, a 27B vision-language model, for llama.cpp, Ollama and LM Studio, with the speculative-decoding head kept. For people running a large model locally. Kvark-Sparkle-1.0-27B-GGUF is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by Egzakta ## About ### What it is This repository holds GGUF files of KVARK Sparkle 1.0, a 27B image-and-text model, for llama.cpp, Ollama and LM Studio. Every file keeps the MTP head, so speculative decoding works. The quantizations use an importance matrix computed from the maker's own SFT data, included as a file. ### Who it suits People who run a large model on their own machine and want the quant file sized to a 24 GB card or less. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Run a local 27B model with an OpenAI-style API | A local chat endpoint | | Research | AI hobbyists | Compare quant sizes and reasoning cost | Q4\_K\_M against Q6\_K | | Education | Students | Try a vision-language model at home | Ask about an image with the mmproj file | | IT | Small teams | Host a private assistant | An internal chat tool through Ollama | You can skip it if your machine cannot hold a file over 15 GB, or if you want a small model. ### Where to use it - llama.cpp, through llama-server - Ollama - LM Studio, where you search for the repository name | Use case | Who | What they get back | | --- | --- | --- | | Local chat | Developers | Text answers, with thinking returned in a separate field | | Image questions | Hobbyists | Answers about an image through the vision projector file | | Faster decoding | Developers | Speculative decoding through the MTP head | ### How to set it up - Python and pip on your computer, to download - llama.cpp, Ollama or LM Studio - Disk and memory for one file, from 15.3 GB to 29.0 GB 1. Install the Hugging Face hub package. 2. Pick a quant. The card recommends Q4\_K\_M for 24 GB cards. 3. Start it in llama.cpp, Ollama or LM Studio. ```sh pip install -U huggingface_hub hf download Egzakta/Kvark-Sparkle-1.0-27B-GGUF ``` ### How to use it Start llama-server with the Q4\_K\_M quant and a 65,536 context, then chat through the OpenAI-style API on port 8080. The server returns the thinking in a field called reasoning\_content. ```bash llama-server -hf Egzakta/Kvark-Sparkle-1.0-27B-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 | | --- | --- | | --spec-type draft-mtp | Switches on faster drafting using the bundled MTP head, with a draft cap of 3 tokens | | --no-mmproj | Text only, skipping the vision projector | | -ctk q8\_0 -ctv q8\_0 | Halves the KV cache | | --reasoning off | Answers without thinking | ```bash curl -s localhost:8080/v1/chat/completions -H "Content-Type: application/json" \ -d '{"messages": [{"role": "user", "content": "How many primes are there between 100 and 150?"}]}' ``` ### Example You ask the server how many primes lie between 100 and 150. It reasons first, returns that thinking in reasoning\_content, and gives the answer in the message. ### For newcomers and experts If you are new, pick Q4\_K\_M, use the llama-server command and ask one question. In LM Studio, the card suggests a 65,536-token context with the same sampling values as the llama-server command. If you know the field, the card measures KV cache by context. At 262,144 tokens it is 16 GiB at f16 and 8.5 GiB at q8\_0, plus under 0.2 GB of linear-attention state. The hero image reports a median answer time of 13.7 seconds against 28.1 for Qwen3.8-27B. KVARK Sparkle 1.0 27B → Importance matrix from own SFT data → GGUF quants with the MTP head → llama.cpp, Ollama or LM Studio Median time to answer in seconds | Label | Value | | --- | --- | | Qwen3.8-27B | 28.1 s | | KVARK Sparkle 1.0 | 13.7 s | ### Good to know - License: Apache 2.0 - Cost to run: your own hardware, with no paid service named - Data it reaches: the files you download and the prompts you give it - Memory: Q4\_K\_M is 16.8 GB, and the KV cache adds more as context grows - Images: need the separate mmproj file ## README ````markdown # Kvark-Sparkle-1.0-27B-GGUF GGUF files of KVARK Sparkle 1.0, a 27B vision-language model, for llama.cpp, Ollama and LM Studio, with the speculative-decoding head kept. For people running a large model locally. 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 1.0, a 27B image-and-text model, for llama.cpp, Ollama and LM Studio. Every file keeps the MTP head, so speculative decoding works. The quantizations use an importance matrix computed from the maker's own SFT data, included as a file. Who it suits: People who run a large model on their own machine and want the quant file sized to a 24 GB card or less. ```table Industry | Who uses it | How they use it | Example Software | Developers | Run a local 27B model with an OpenAI-style API | A local chat endpoint Research | AI hobbyists | Compare quant sizes and reasoning cost | Q4_K_M against Q6_K Education | Students | Try a vision-language model at home | Ask about an image with the mmproj file IT | Small teams | Host a private assistant | An internal chat tool through Ollama ``` You can skip it if your machine cannot hold a file over 15 GB, or if you want a small model. Where to use it: - llama.cpp, through llama-server - Ollama - LM Studio, where you search for the repository name ```table Use case | Who | What they get back Local chat | Developers | Text answers, with thinking returned in a separate field Image questions | Hobbyists | Answers about an image through the vision projector file Faster decoding | Developers | Speculative decoding through the MTP head ``` How to set it up: - Python and pip on your computer, to download - llama.cpp, Ollama or LM Studio - Disk and memory for one file, from 15.3 GB to 29.0 GB 1. Install the Hugging Face hub package. 2. Pick a quant. The card recommends Q4_K_M for 24 GB cards. 3. Start it in llama.cpp, Ollama or LM Studio. ```sh pip install -U huggingface_hub hf download Egzakta/Kvark-Sparkle-1.0-27B-GGUF ``` How to use it: Start llama-server with the Q4_K_M quant and a 65,536 context, then chat through the OpenAI-style API on port 8080. The server returns the thinking in a field called reasoning_content. ```bash llama-server -hf Egzakta/Kvark-Sparkle-1.0-27B-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 ``` ```table Tool | What it does --spec-type draft-mtp | Switches on faster drafting using the bundled MTP head, with a draft cap of 3 tokens --no-mmproj | Text only, skipping the vision projector -ctk q8_0 -ctv q8_0 | Halves the KV cache --reasoning off | Answers without thinking ``` ```bash curl -s localhost:8080/v1/chat/completions -H "Content-Type: application/json" \ -d '{"messages": [{"role": "user", "content": "How many primes are there between 100 and 150?"}]}' ``` Example: You ask the server how many primes lie between 100 and 150. It reasons first, returns that thinking in reasoning_content, and gives the answer in the message. For newcomers and experts: If you are new, pick Q4_K_M, use the llama-server command and ask one question. In LM Studio, the card suggests a 65,536-token context with the same sampling values as the llama-server command. If you know the field, the card measures KV cache by context. At 262,144 tokens it is 16 GiB at f16 and 8.5 GiB at q8_0, plus under 0.2 GB of linear-attention state. The hero image reports a median answer time of 13.7 seconds against 28.1 for Qwen3.8-27B. ```flow KVARK Sparkle 1.0 27B Importance matrix from own SFT data GGUF quants with the MTP head llama.cpp, Ollama or LM Studio ``` ```chart Title: Median time to answer in seconds Qwen3.8-27B: 28.1 s KVARK Sparkle 1.0: 13.7 s ``` Good to know: - License: Apache 2.0 - Cost to run: your own hardware, with no paid service named - Data it reaches: the files you download and the prompts you give it - Memory: Q4_K_M is 16.8 GB, and the KV cache adds more as context grows - Images: need the separate mmproj file ## 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: 7,317 - Likes: 14 - License: apache-2.0 - Languages: en - Based on: Egzakta/Kvark-Sparkle-1.0-27B - Tags: gguf, llama.cpp, imatrix, reasoning, kvark, image-text-to-text, conversational ## Links - Model: https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-GGUF - Files: https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-GGUF/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download Egzakta/Kvark-Sparkle-1.0-27B-GGUF ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [Egzakta/Kvark-Sparkle-1.0-27B-GGUF on Hugging Face](https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-GGUF) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Kvark-Sparkle-1.0-27B-GGUF (this page) | 7,317 | 14 | 27.3B | apache-2.0 | 2026-11-05 | | [KVARK Sparkle 1.1 27B GGUF](https://mindket.com/p/e9137ce1-b518-4683-b206-1eea9d1e3e12) | 610 | 3 | 27.3B | apache-2.0 | 2026-11-10 | | [Thinklite-27B-GGUF](https://mindket.com/p/7943692c-1782-4340-9958-f2c39d245585) | 1,796 | 18 | 26.9B | apache-2.0 | 2026-11-04 | | [KVARK Sparkle 1.0 (27B) FP8](https://mindket.com/p/acc7b74d-4609-4c56-abe3-47d4ccc946e3) | 24,270 | 5 | 27.8B | apache-2.0 | 2026-10-29 | | [Kvark-Sparkle-1.0-27B](https://mindket.com/p/6451d52d-00e1-4878-8a25-80910fc67ac7) | 315 | 15 | 27.4B | apache-2.0 | 2026-11-05 | | [KVARK Sparkle 1.1 (27B)](https://mindket.com/p/fb8f86c3-be94-494c-9f8e-66d3faba633c) | 59 | 3 | 27.4B | apache-2.0 | 2026-11-10 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-11-05 - First published: 2026-10-19 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Egzakta/Kvark-Sparkle-1.0-27B-GGUF on Hugging Face](https://huggingface.co/Egzakta/Kvark-Sparkle-1.0-27B-GGUF) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/9198b0c1-c121-4820-9f9b-28e4e7d71968) - [Its page on Mindket](https://mindket.com/p/9198b0c1-c121-4820-9f9b-28e4e7d71968)