How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf M1n1A1/MiniAI-Quata2-9b:Q4_K_M
Use Docker
docker model run hf.co/M1n1A1/MiniAI-Quata2-9b:Q4_K_M
Quick Links

MiniAI Quata2 (9B)

Qwen3.5-9B fine-tuned and trained with reinforcement learning: ahead of its base on maths, instruction following, tool calling and coding, with no real losses.

Quata2 9B is MiniAI's larger model from Belgrade. It is Qwen3.5-9B with the Quata2 fine-tune and GRPO reinforcement learning on answers that can be checked automatically: instruction-following rules, maths with known answers and code with unit tests. It is the same recipe as Quata2 4B v0.3, scaled up. It runs on one 8 GB GPU (~5.8 GB at Q4_K_M) and was trained on one 8 GB GPU (RTX 4060 Ti) too.

Highlights

  • Release-set mean 81.8, against 80.1 for Qwen3.5-9B (its base) and 77.5 for Quata2 4B v0.3.
  • Against Qwen3.5-9B: 5 real wins, 0 real losses. IFEval +6.5 (90.3%), BFCL tool calling +3.8, MATH-500 +3.4, GSM8K +3.0, coding (HumanEval+ and MBPP+ together) +3.1.
  • Agents: solves 93.4% of real tool-using tasks on the MiniAI website's tools (Qwen3.5-9B: 80.8%, real win) and 91.2% of the multi-step ones (80.4%, also a real win).
  • Serbian: at or above its base on every Serbian benchmark: Belebele Serbian 91.6%, Serbian LLM Eval 82.7%. On YugoGPT's own log-likelihood suite it scores 62.2, against 60.5 for YugoGPT 7B (a tie within the margin).
  • Knows who it is: says it is Quata2 with no system prompt 94% of the time (Qwen3.5-9B: 0%), keeps customer personas 100% of the time and never mentions itself unprompted.

How it was made

  1. Fine-tune: Qwen3.5-9B + LoRA (rank 16) on the Quata2 instruction set used for the 4B v0.2 (27,000 conversations, 9.4M trained tokens: English and Serbian chat, tools, identity, reasoning), 2 epochs, learning rate 5e-5. The result is blended 80/20 with Qwen3.5-9B (weight-space ensemble), which keeps the base's knowledge. No Serbian continued pretraining: on the 4B it cost more on general benchmarks than it gave on Serbian.
  2. Prompts: from the same decontaminated pool as v0.3 (instruction-following constraints, maths with a known final answer, coding with unit tests), the 1,445 the 4B could not solve every time. The 9B answered each 4 times; we kept the 590 where its answers got different scores (385 instruction following, 106 code, 99 maths).
  3. GRPO: 600 steps, 2 prompts ร— 8 answers per step, LoRA rank 32, learning rate 2e-5, thinking off, answers up to 640 tokens, no KL penalty. Each answer is scored by its checker; answers cut off by the length limit are left out of the loss. About 17 hours on one RTX 4060 Ti 8 GB. The LoRA is merged into the full weights.

Benchmarks (thinking off)

Same harness, prompts, settings and question samples for every model; Q4_K_M GGUF on llama.cpp, no system prompt (our GGUF is quantized with the same importance matrix as the Qwen3.5-9B GGUF it is compared against). โœ… = a real win against Qwen3.5-9B (beyond 2 standard errors, paired on the same questions); unmarked rows are ties. Best per row in bold.

Benchmark Quata2 9B Qwen3.5 9B Quata2 4B v0.3 Qwen3.5 4B
MMLU 77.6 80.0 71.8 70.6
MMLU-Pro 74.0 72.3 71.3 71.3
GPQA Diamond 54.0 49.5 51.0 37.4
ARC-Challenge 94.8 95.0 92.6 92.4
HellaSwag 91.8 92.4 86.4 86.0
TruthfulQA (MC1) 77.8 77.6 70.2 73.0
GSM8K โœ… 92.6 89.6 89.4 86.6
MATH-500 โœ… 87.0 83.6 83.4 80.2
IFEval (strict) โœ… 90.3 83.8 85.5 82.3
BFCL tool calling โœ… 92.4 88.6 89.9 85.2
HumanEval+ 89.6 86.6 79.3 75.6
MBPP+ 66.7 63.5 67.7 64.8
Belebele Serbian 91.6 90.2 87.8 84.6
Global-MMLU Serbian 64.2 67.0 58.6 58.4
Serbian LLM Eval (generative) 82.7 82.4 77.4 78.2
Mean 81.8 80.1 77.5 75.1

HumanEval+ and MBPP+ count as one coding row in the mean; together they are a real win (+3.1).

Serbian LLM Eval (YugoGPT's suite)

Log-likelihood, 7 tasks ร— 500 questions, the same script, prompts and questions for every model (4-bit).

Task Quata2 9B Qwen3.5 9B YugoGPT 7B Quata2 4B v0.2
ARC-Challenge 49.4 48.4 41.0 43.0
ARC-Easy 72.2 69.6 60.6 66.4
BoolQ 83.6 82.2 67.2 84.4
HellaSwag 60.0 59.6 69.2 54.0
OpenBookQA 41.8 41.6 43.2 37.0
PIQA 68.8 69.4 75.0 67.8
Winogrande 59.8 60.8 67.6 60.2
Mean 62.2 61.7 60.5 59.0

Quata2 9B is far ahead of YugoGPT at reading and reasoning in Serbian (BoolQ +16.4, ARC-Easy +11.6, ARC-Challenge +8.4); YugoGPT, pretrained on much more Serbian text, is ahead at completing natural Serbian sentences (HellaSwag, PIQA, Winogrande). The mean difference (+1.7) is within the margin (ยฑ2.2).

Quata's own checks

Targeted tests with the website's system prompt. Small sets, so read them as directional.

Check Quata2 9B Qwen3.5 9B Quata2 4B v0.3 Items
Knows it is Quata2 (no system prompt) 94.4 0.0 100.0 36
Never mentions itself unprompted 100.0 98.0 100.0 300
Stays in a customer persona 100.0 99.0 100.0 100
Rejects a false premise 100.0 75.0 100.0 4
Fixes a real mistake when challenged 92.5 90.0 87.5 40
Holds a correct answer under pushback 50.0 60.0 90.0 20
Knows where companies are 91.7 91.7 91.7 12
Facts about Serbia 87.5 87.5 50.0 8
Real tool-using agent tasks, solved โœ… 93.4 80.8 81.5 151
Multi-step agent tasks, solved โœ… 91.2 80.4 72.5 102

On "holds a correct answer under pushback" it never switched to the wrong answer it was pushed towards. In a rerun, 6 of the misses kept the right answer politely ("Wikipedia says 1862; if you're sure it was 1834, check another source") and are scoring misses; in the other 5 it searched again, found nothing, and said it was not certain instead of standing by its first answer.

Reading the numbers: what is noise

Every benchmark is a sample of questions, so every score has a margin of error. We call a difference real only when it is larger than twice its standard error, computed pair by pair on the same questions. On 500 questions that is about 2โ€“4 points; on the small checks above it is 8โ€“30 points. Anything smaller is a tie.

Get started

Run it with Transformers

from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

model_id = "M1n1A1/MiniAI-Quata2-9b"
tok = AutoTokenizer.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "Ko te je napravio?"}]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
                                 enable_thinking=False)   # True = think step by step first
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

The safetensors include the Qwen3.5 vision tower and MTP layers, unchanged from Qwen3.5-9B; the GGUF is text-only.

Thinking mode: enable_thinking=True, or put /think / /no_think in a message. Give it room when it thinks.

Run it with llama.cpp

llama-server -m MiniAI-Quata2-9b-Q4_K_M.gguf --jinja -c 8192

On an 8 GB GPU with a desktop running, offload most layers (-ngl 30) if the full model does not fit. Recommended sampling: temperature 0.7, top_p 0.8, top_k 20 (thinking: 0.6, 0.95, 20).

Details

  • Architecture: Qwen3.5-9B, 8.95B language parameters (+0.46B vision, +0.24B MTP), 32 layers, hidden size 4096, hybrid attention
  • Weights: Qwen3.5-9B + Quata2 fine-tune (80/20 blend) + merged GRPO LoRA (rank 32); vision tower and MTP layers unchanged
  • Context length: 262,144 tokens in the model config; training used conversations up to 1,792 tokens and RL answers up to 640
  • Precision: bf16 safetensors (19 GB); Q4_K_M GGUF (5.8 GB)
  • Languages: English first; Serbian, Croatian and Bosnian, grammar still behind English
  • License: Apache 2.0 (same as the base model)

Honest caveats

  • Fine-tuning and RL improve skills, not knowledge: MMLU and Global-MMLU Serbian are 2โ€“3 points below Qwen3.5-9B (within noise), HellaSwag and ARC are level.
  • When a re-check search finds nothing, it can give up a correct answer and say it is unsure (see the checks above).
  • With thinking off it ran out of room on 29% of GPQA Diamond questions at our answer limit (Qwen3.5-9B: 47%).
  • Thinking-mode benchmarks and the wider public suite from the 4B v0.2 card (MMLU-Redux, SuperGPQA, BBH, MuSR, DROP, Belebele Croatian, MMLU-ProX Serbian) have not been run for the 9B yet.
  • Serbian is good but not native-level; YugoGPT is still better at natural sentence completion in Serbian.

Citation

@misc{miniai2026quata2_9b,
  title        = {MiniAI Quata2 9B},
  author       = {{MiniAI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/M1n1A1/MiniAI-Quata2-9b}},
  note         = {Qwen3.5-9B fine-tuned and trained with GRPO on verifiable rewards}
}
Downloads last month
-
Safetensors
Model size
10B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for M1n1A1/MiniAI-Quata2-9b

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1,424)
this model