KVARK Sparkle 1.1 (27B)

KVARK Sparkle 1.1 (27B)

A 27B model tuned for document work: it searches a document base, answers from what it finds and says so when the answer is missing, for builders of document and search agents.

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 1.1 is a 27B model from Egzakta, trained from Sparkle 1.0 for document work. Given a pile of documents, it looks things up, builds its reply from the hits and owns up when nothing answers the question. It keeps the Qwen3.8-27B architecture and takes images and text in. On public benchmarks the card reports no significant difference from Qwen3.8-27B.

Who it suits: Developers who build document question-answering or search agents will get the most from it. It suits cases where a wrong answer is worse than no answer.

Industry | Who uses it | How they use it | Example
Legal | Paralegals and legal-tech builders | Search a pile of contracts and answer only from them | Finding a clause, or learning it is absent
Compliance | Analysts | Check policy documents for a specific requirement | Confirming that a rule is or is not covered
Customer support | Support engineers | Answer from a help-centre document base | A reply that says the manual does not cover the question
Research | Knowledge workers | Query a private set of papers or reports | Pulling one fact out of a long report

If you do not build agents that search documents, you can skip it, since its tuning targets that job.

Where to use it:

• Python with the transformers library, which Hugging Face lists
• vLLM, using the separate FP8 repository
• llama.cpp, using the separate GGUF repository
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Document Q&A | Agent builders | Answers drawn from the documents you supply
Unanswerable questions | Compliance teams | A clear NOT IN DOCUMENTS reply when the fact is missing
Reasoning tasks | Developers | Answers that use fewer reasoning tokens than the base model

How to set it up: You need Python with pip, a GPU with room for the weights, and a search tool your agent can call. The BF16 weights are about 54.7 GB. The card also names smaller variants.

• Python with pip, on Windows, macOS or Linux
• The transformers library, or vLLM or llama.cpp for the other variants
• A GPU with enough memory, about 54.7 GB for BF16
• Your own agent with search tools and a system prompt

1. Install the Hugging Face hub package.
2. Download the model files.
3. Add the stop rule from stop_system.txt to the end of your system prompt.

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

Variant | Size | Use
BF16 with MTP head | 55.6 GB | This repository
FP8 with MTP head | 30.9 GB | For vLLM, in the Kvark-Sparkle-1.1-27B-FP8 repository
GGUF, BF16 to IQ4_XS | 15.3 to 54.7 GB | For llama.cpp, in the Kvark-Sparkle-1.1-27B-GGUF repository

How to use it: Put the stop rule after your tool instructions in the system prompt. The rule tells the model that three searches in a row with no page stating the fact mean the documents likely lack it. It should then reply NOT IN DOCUMENTS and always end with a Final answer line. The model was trained with this rule.

from huggingface_hub import hf_hub_download

stop_rule = open(hf_hub_download("Egzakta/Kvark-Sparkle-1.1-27B", "stop_system.txt")).read().strip()
system = tool_instructions + "\n\n" + stop_rule  # tool_instructions: your agent's own system prompt

Example: Your agent searches a document base for a fact. After three searches in a row, none of the pages found has that fact. The model stops searching and answers NOT IN DOCUMENTS instead of guessing.

For newcomers and experts: An agent is a program that lets a model call tools, such as a document search. A good first step is to load the stop rule into your system prompt and ask a question you know the documents do not answer.

The model keeps the Qwen3.8-27B design of 64 layers, 48 with linear attention and 16 with full attention, plus a one-layer MTP head for speculative decoding. Context is 262,144 tokens natively, evaluated up to 64k. The card says the median answer uses 7 to 48 percent fewer reasoning tokens than Qwen3.8-27B. Its document numbers all use the stop rule.

You add the stop rule to your system prompt
The model searches your document base
It answers from the pages it finds
Or it replies NOT IN DOCUMENTS and stops

Title: Questions with no answer in the documents, correctly declined
Sparkle 1.0: 36 percent
Sparkle 1.0 with the stop rule: 74 percent
Sparkle 1.1 with the stop rule: 91.5 percent

Good to know:

• License: apache-2.0, as listed on Hugging Face
• Cost to run: none named, apart from a GPU for the weights
• Data it reaches: the documents and tools you connect to your agent, and the source names no other access
• Languages listed are English and Serbian
• Not for: document agents run without the stop rule, since the model then declines far fewer unanswerable questions
• The card's document-work scores come from the authors' own evaluation

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: 59
• Likes: 3
• License: apache-2.0
• Languages: en, sr
• Based on: Egzakta/Kvark-Sparkle-1.0-27B
• Quantized versions: 4
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, reasoning, efficient-reasoning, agentic-rag, document-qa, qwen3.8, kvark, conversational, model-index

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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