Benchmarks (raw numbers, no rounding in our favour)
All three models were scored the same way: our own harness, Q4_K_M GGUF on llama-server, same prompts and
settings. Higher is better.
Table with columns: Test, Preview, Qwen3.5-4B (base), Quata2 4B v0.1 (Qwen3-4B base)| Test | Preview | Qwen3.5-4B (base) | Quata2 4B v0.1 (Qwen3-4B base) |
|---|
| BFCL tool calling, overall | 87.9 | 85.1 | 89.8 |
| MMLU (knowledge) | 72.6 | 70.2 | 65.0 |
| GSM8K (grade-school maths) | 85.2 | 87.0 | 84.2 |
| IFEval, prompt-level strict | 74.6 | 81.5 | 76.0 |
| HumanEval+ (coding) | 65.9 | 78.0 | — |
| MBPP+ (coding) | 58.7 | 64.6 | — |
| Agent tasks on real tools, solved | 80.1 | 80.1 | 56.3 |
| Agent tasks, multi-step | 70.6 | 74.5 | 46.1 |
| Double-check: keeps a right answer when challenged (20 items) | 80.0 | 65.0 | 25.0 |
| Double-check: fixes a wrong answer when challenged (40 items) | 82.5 | 95.0 | 62.5 |
| False-premise questions handled (4 items) | 75.0 | 100 | 75.0 |
| Says it is Quata2 with no system prompt | 97.2 | 0.0 | 100 |
Where it stands: ahead of the base on tool calling, knowledge and standing firm; behind it on IFEval,
coding, multi-step agent tasks and false premises. Rows marked with item counts use tiny test sets (the false-premise
gap is a single question), so treat them as hints; IFEval (524 prompts) and coding (164 and 378 problems) are
solid, and those gaps are real.
Run it
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration model_id = "M1n1A1/MiniAI-Quata2-4b-Preview"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)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; the GGUF is text-only.
Use Qwen3_5ForConditionalGeneration (not AutoModelForCausalLM) with these files.
Ollama (straight from Hugging Face):
ollama run hf.co/M1n1A1/MiniAI-Quata2-4b-Preview
llama.cpp:
llama-server -m MiniAI-Quata2-4b-Preview-Q4_K_M.gguf --jinja -c 8192
The GGUF carries Qwen3.5's chat template (with native tool calling and thinking on/off), so use --jinja.
Recommended sampling: temperature 0.7, top_p 0.8, top_k 20. Text only: the vision part of Qwen3.5 is not
included in this file.
Details
- Base: Qwen/Qwen3.5-4B
- Training: supervised fine-tune on OpenHermes-2.5, Tulu-3 instruction following, Hermes function calling,
ToolACE and When2Call, plus MiniAI-made identity, tool error-recovery and double-check data; a tool-use
fine-tune merged in at 30%; then the weights blended 80/20 toward Qwen3.5-4B (WiSE-FT) to keep the base
model's strengths.
- GGUF:
MiniAI-Quata2-4b-Preview-Q4_K_M.gguf (2.8 GB, 4-bit), sha256
6e8d84f020dc971382e6c04d3b2c3caec363af7a235fddc9457b4b953b8aaecd
- Safetensors:
model-00001-of-00002.safetensors (~5.9 GB), model-00002-of-00002.safetensors
(~3.9 GB), model-mtp.safetensors (~0.24 GB), and model.safetensors.index.json; bf16,
text + vision.
- Languages: English first; Serbian, Croatian and Bosnian work but grammar still lags behind English.
- Known weaknesses: strict formatting instructions, coding, and occasionally confident wrong facts. Check
important answers.
- License: Apache 2.0 (same as the base model).
Citation
@misc{miniai2026quata2preview, title = {MiniAI Quata2 4B Preview}, author = {{MiniAI}}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/M1n1A1/MiniAI-Quata2-4b-Preview}}, note = {Preview build, fine-tuned from Qwen3.5-4B}}