Variants
Use it with the stop rule
For document and search agents, add stop_system.txt (in this repository) to the end of your system prompt, after your tool instructions:
Searching has a cost. If three searches in a row bring back no page that states the specific fact asked for, the documents most likely do not contain it: stop searching and reply with Final answer: NOT IN DOCUMENTS. Never end a reply without a Final answer line.
The model was trained with this rule, and every document-work number on this card uses it. Without it, the model declines far fewer questions whose answer is not in the documents.
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
Model details
Document work
GroundedWork is our agentic RAG benchmark: the model gets search and read_page tools over public Serbian documents and answers factual, chronology, finance and cross-document questions. Every run uses the same harness and tools with thinking on, and DeepSeek V4.1 Flash judges the answers against gold answers. Each question runs three times; the scores are over all runs.
The stop rule alone takes Sparkle 1.0 from 86.2 to 89.5 and from 36% to 74%. Against Sparkle 1.0 with the same rule, Sparkle 1.1 adds 1.5 points on document Q&A and 17.5 points on declining.
Results
Same harness for both models: FP8 weights, thinking on, a 64k token cap, the same prompts and sampling. Differences are paired over items, with 95% bootstrap intervals, and computed before rounding. FRAMES answers are judged by two LLMs, GLM-5.3-Flash and MiMo-V2.6-Flash; when they disagree, a rule-based check decides.
No difference is statistically significant. Where Sparkle 1.1 scores lower (AIME, GPQA Diamond, FRAMES), the difference is inside the interval.
General ability vs Sparkle 1.0
A held-out split, paired on the same items, with FRAMES judged the same way. No difference is statistically significant.
Speed
Measured with vLLM 0.32.0, FP8 weights and 1 MTP draft token for both models, on the 20 fixed questions of the Sparkle 1.0 card (10 GPQA Diamond, 10 LiveCodeBench). Requests ran one at a time with thinking on, a 32k token cap, temperature 0.6, top_p 0.95 and top_k 20. The median is the 11th of the 20 sorted times, as on the 1.0 card. The GPU, vLLM version and MTP setting differ from the 1.0 card's speed test, so its times are not comparable with these.
Quickstart
vLLM (FP8 with MTP)
vllm serve Egzakta/Kvark-Sparkle-1.1-27B-FP8 \ --max-model-len 131072 --kv-cache-dtype fp8 \ --reasoning-parser qwen3 \ --enable-auto-tool-choice --tool-call-parser qwen3_coder \ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")response = client.chat.completions.create( model="Egzakta/Kvark-Sparkle-1.1-27B-FP8", messages=[{"role": "user", "content": "How many primes are there between 100 and 150?"}], max_tokens=65536,)message = response.choices[0].messageprint(getattr(message, "reasoning", None))print(message.content)print(response.usage.completion_tokens, "tokens")
llama.cpp
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
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor model_id = "Egzakta/Kvark-Sparkle-1.1-27B"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto") messages = [{"role": "user", "content": [{"type": "text", "text": "How many primes are there between 100 and 150?"}]}]inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)output = model.generate(**inputs, max_new_tokens=65536)print(processor.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
Thinking and sampling
Thinking is on by default. To turn it off for a request, pass extra_body={"chat_template_kwargs": {"enable_thinking": False}}. Use temperature 1.0, top_p 0.95 and top_k 20, the defaults in generation_config.json. For hard math and code, allow max_tokens of 65,536, the cap used in our evaluation.
For the hardest problems, a hard thinking cap can help: with vLLM's per-request thinking_token_budget at 22,528 tokens and a 32k answer cap, Sparkle 1.0 scored higher than with uncapped thinking on our held-out split (LiveCodeBench 82.9 vs 73.7, AIME 87.5 vs 83.3, GPQA Diamond 86.6 vs 83.5). Without a cap, long traces can run into the answer limit. This was measured on Sparkle 1.0.
How it was trained
LoRA (rank 64) on Sparkle 1.0, multi-turn supervised fine-tuning, merged into the weights:
- Agent trajectories over a public Serbian legal corpus (Legislation RS, CC BY). Questions were generated by DeepSeek V4.1 Flash (MIT license), and the answers and tool use come from Sparkle 1.0, checked against gold answers. For questions with no answer in the corpus, trajectories are cut after two or three search rounds and end with the decline.
- Multi-hop document QA from MuSiQue (CC BY 4.0), answered by Sparkle 1.0, keeping the shortest correct answer.
- Retention data generated by Sparkle 1.0 itself (instruction following, multiple choice, math, code), checked by verifiers.
No outputs of proprietary models were used. Evaluation questions were never used for training.
Limitations
- English questions over Serbian documents are the weakest case (cross-lingual retrieval).
- Period traps: a question about a period that is not in the documents (for example a rate for a future month) can be answered with the nearest available value instead of declined.
- Questions that need links between several documents are the weakest GroundedWork family.
- Planning and scheduling did not improve over Sparkle 1.0.
- Vision uses the base model's encoder and was not separately evaluated. Context beyond 64k tokens was not measured.
- No separate safety evaluation was run; we did not train for or against refusals.
License
Apache 2.0, inherited from Qwen3.8-27B. See LICENSE and NOTICE. Commercial use is permitted, with no revenue cap.
hello@kvark.ai
Citation
@misc{kvarksparkle2026_11, title = {KVARK Sparkle 1.1}, author = {Egzakta AI Lab}, year = {2026}, url = {https://huggingface.co/Egzakta/Kvark-Sparkle-1.1-27B}}