M1n1A1

MiniAI-Quata2-4b

README

License: apache-2.0

Highlights

Beats Qwen3-4B — wins 7 of 8 knowledge and common-sense benchmarks against its own base
Native thinking mode — Qwen3 <think> reasoning, switch it on or off per message
Honest under pressure — holds correct answers when pushed back on, owns real mistakes
Knows the Balkans — better Serbian facts, no more putting every company in Belgrade
Speaks BCMS — tops the Serbian LLM Eval among the 4B models we tested
100% on-device option — nothing leaves your machine

Benchmarks

Every number below was measured on the same harness, with the same questions, for all three models: 500 questions per benchmark, zero-shot.

ARC-Challenge — hard grade-school science

ARC-Easy — grade-school science

HellaSwag — common-sense sentence completion

Winogrande — pronoun and common-sense reasoning

PIQA — physical common sense PIQA

BoolQ — yes/no reading comprehension BoolQ

OpenBookQA — science facts plus reasoning

TruthfulQA — avoiding common misconceptions

IFEval — following precise instructions IFEval

GSM8K — grade-school maths GSM8K

MMLU — knowledge across 57 subjects MMLU

Serbian LLM Eval — the same kind of tasks, in Serbian

Table with columns: Benchmark, MiniAI Quata2 (4B), Qwen3-4B (base), MiniAI Quata1.5 (4B)
Benchmark	MiniAI Quata2 (4B)	Qwen3-4B (base)	MiniAI Quata1.5 (4B)
ARC-Challenge	54.8%	50.6%	52.4%
ARC-Easy	82.4%	78.4%	78.2%
HellaSwag	70.8%	67.2%	69.0%
Winogrande	71.8%	67.6%

Quata2 leads on 10 of 12 benchmarks, beats its own Qwen3-4B base on the knowledge and common-sense suite by +2.4 points on average, and jumps +8.2 points on GSM8K maths over Quata1.5. That's the return you get from fixing a model's flaws instead of papering over them.

The fixes, measured

Quata2 was built specifically to fix Quata1.5's weak spots. Our own targeted checks (small test sets, so read them as directional):

Table with columns: Check, Quata1.5, Quata2
Check	Quata1.5	Quata2
Doesn't put every company in Belgrade	45%	100%
Knows where companies are	58%	83%
Facts about Serbia	50%	62%
Holds correct answers under pushback, owns real mistakes	67%	100%
Knows who it is, never claims to be ChatGPT	90%	100%
Thinking mode reaches the right answer

Get started

Run it locally (Ollama)

bash

ollama run miniai/quata2:4b

Model page: ollama.com/MiniAI/quata2

Thinking is on by default; turn it off for quick replies with /set nothink in the chat, or "think": false in the API. The Ollama build is the Q4_K_M GGUF (~2.5 GB) with Qwen3's hybrid-thinking chat template.

Run it with Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "M1n1A1/MiniAI-Quata2-4b"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.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 firstinputs = 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))

Thinking mode: enable_thinking=True, or put /think / /no_think in a message. Give it room (2,000–4,000 new tokens) when it thinks.

Hosted API

Soon™

Details

Architecture: Qwen3-based, 4B parameters, 36 layers, hidden size 2560
Context length: 40,960 tokens
Precision: bfloat16 safetensors (~8 GB); Q4_K_M GGUF (~2.5 GB) on Ollama
Training: supervised fine-tune on ~108k curated conversations (OpenHermes-2.5, Tulu-3 instruction following, MiniAI identity data, self-distilled verified reasoning) → preference tuning (DPO) on Quata1.5's failure cases → weight blend toward Qwen3-4B (WiSE-FT, λ = 0.8) to keep the base model's strengths
Languages: English first; Serbian, Croatian and Bosnian understood and spoken, with grammar still behind English — better Balkan-language ability is the focus of the next Quata models
License: Apache 2.0 (base)

Citation

If you use Quata2 in your work, please cite:

bibtex

@misc{miniai2026quata2,  title        = {MiniAI Quata2 (4B)},  author       = {{MiniAI}},  year         = {2026},  publisher    = {Hugging Face},  howpublished = {\url{https://huggingface.co/M1n1A1/MiniAI-Quata2-4b}},  note         = {Fine-tuned from Qwen3-4B. Ollama: \url{https://ollama.com/MiniAI/quata2}}}

and the base model:

bibtex

@misc{qwen3technicalreport,  title         = {Qwen3 Technical Report},  author        = {{Qwen Team}},  year          = {2025},  eprint        = {2505.09388},  archivePrefix = {arXiv},  primaryClass  = {cs.CL},  url           = {https://arxiv.org/abs/2505.09388}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

M1n1A1

Model Tree

Base

Qwen/Qwen3-4B

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Highlights

Beats Qwen3-4B — wins 7 of 8 knowledge and common-sense benchmarks against its own base
Native thinking mode — Qwen3 <think> reasoning, switch it on or off per message
Honest under pressure — holds correct answers when pushed back on, owns real mistakes
Knows the Balkans — better Serbian facts, no more putting every company in Belgrade
Speaks BCMS — tops the Serbian LLM Eval among the 4B models we tested
100% on-device option — nothing leaves your machine

Benchmarks

Every number below was measured on the same harness, with the same questions, for all three models: 500 questions per benchmark, zero-shot.

ARC-Challenge — hard grade-school science

ARC-Easy — grade-school science

HellaSwag — common-sense sentence completion

Winogrande — pronoun and common-sense reasoning

PIQA — physical common sense PIQA

BoolQ — yes/no reading comprehension BoolQ

OpenBookQA — science facts plus reasoning

TruthfulQA — avoiding common misconceptions

IFEval — following precise instructions IFEval

GSM8K — grade-school maths GSM8K

MMLU — knowledge across 57 subjects MMLU

Serbian LLM Eval — the same kind of tasks, in Serbian

Table with columns: Benchmark, MiniAI Quata2 (4B), Qwen3-4B (base), MiniAI Quata1.5 (4B)
Benchmark	MiniAI Quata2 (4B)	Qwen3-4B (base)	MiniAI Quata1.5 (4B)
ARC-Challenge	54.8%	50.6%	52.4%
ARC-Easy	82.4%	78.4%	78.2%
HellaSwag	70.8%	67.2%	69.0%
Winogrande	71.8%	67.6%

The fixes, measured

Quata2 was built specifically to fix Quata1.5's weak spots. Our own targeted checks (small test sets, so read them as directional):

Table with columns: Check, Quata1.5, Quata2
Check	Quata1.5	Quata2
Doesn't put every company in Belgrade	45%	100%
Knows where companies are	58%	83%
Facts about Serbia	50%	62%
Holds correct answers under pushback, owns real mistakes	67%	100%
Knows who it is, never claims to be ChatGPT	90%	100%
Thinking mode reaches the right answer

Get started

Run it locally (Ollama)

bash

ollama run miniai/quata2:4b

Model page: ollama.com/MiniAI/quata2

Run it with Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "M1n1A1/MiniAI-Quata2-4b"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.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 firstinputs = 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))

Thinking mode: enable_thinking=True, or put /think / /no_think in a message. Give it room (2,000–4,000 new tokens) when it thinks.

Hosted API

Soon™

Details

Architecture: Qwen3-based, 4B parameters, 36 layers, hidden size 2560
Context length: 40,960 tokens
Precision: bfloat16 safetensors (~8 GB); Q4_K_M GGUF (~2.5 GB) on Ollama
Training: supervised fine-tune on ~108k curated conversations (OpenHermes-2.5, Tulu-3 instruction following, MiniAI identity data, self-distilled verified reasoning) → preference tuning (DPO) on Quata1.5's failure cases → weight blend toward Qwen3-4B (WiSE-FT, λ = 0.8) to keep the base model's strengths
Languages: English first; Serbian, Croatian and Bosnian understood and spoken, with grammar still behind English — better Balkan-language ability is the focus of the next Quata models
License: Apache 2.0 (base)

Citation

If you use Quata2 in your work, please cite:

bibtex

@misc{miniai2026quata2,  title        = {MiniAI Quata2 (4B)},  author       = {{MiniAI}},  year         = {2026},  publisher    = {Hugging Face},  howpublished = {\url{https://huggingface.co/M1n1A1/MiniAI-Quata2-4b}},  note         = {Fine-tuned from Qwen3-4B. Ollama: \url{https://ollama.com/MiniAI/quata2}}}

and the base model:

bibtex

@misc{qwen3technicalreport,  title         = {Qwen3 Technical Report},  author        = {{Qwen Team}},  year          = {2025},  eprint        = {2505.09388},  archivePrefix = {arXiv},  primaryClass  = {cs.CL},  url           = {https://arxiv.org/abs/2505.09388}}