Instructions to use Acnoryx/Airy-Core-0.6B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Acnoryx/Airy-Core-0.6B with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Acnoryx/Airy-Core-0.6B",
	filename="gguf/acnoryx-0.6b-f16.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use Acnoryx/Airy-Core-0.6B with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Acnoryx/Airy-Core-0.6B:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Acnoryx/Airy-Core-0.6B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Acnoryx/Airy-Core-0.6B:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Acnoryx/Airy-Core-0.6B:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Acnoryx/Airy-Core-0.6B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Acnoryx/Airy-Core-0.6B:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Acnoryx/Airy-Core-0.6B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Acnoryx/Airy-Core-0.6B:Q4_K_M

Use Docker

docker model run hf.co/Acnoryx/Airy-Core-0.6B:Q4_K_M

LM Studio
Jan
Ollama
How to use Acnoryx/Airy-Core-0.6B with Ollama:
```
ollama run hf.co/Acnoryx/Airy-Core-0.6B:Q4_K_M
```

Unsloth Studio new

How to use Acnoryx/Airy-Core-0.6B with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Acnoryx/Airy-Core-0.6B to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Acnoryx/Airy-Core-0.6B to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Acnoryx/Airy-Core-0.6B to start chatting

Docker Model Runner
How to use Acnoryx/Airy-Core-0.6B with Docker Model Runner:
```
docker model run hf.co/Acnoryx/Airy-Core-0.6B:Q4_K_M
```

Lemonade

How to use Acnoryx/Airy-Core-0.6B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Acnoryx/Airy-Core-0.6B:Q4_K_M

Run and chat with the model

lemonade run user.Airy-Core-0.6B-Q4_K_M

List all available models

lemonade list

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Acnoryx 0.6B — Release GGUF

Acnoryx AI is a dermatology-focused language model fine-tuned for the Acnoryx acne scanner app. It provides skincare guidance, acne analysis, and scan interpretation in Vietnamese and English.


Base model	`Qwen/Qwen3-0.6B`
Method	SFT (Supervised Fine-Tuning) via Unsloth LoRA
Languages	Vietnamese, English
Domain	Acne types, causes, treatments, skincare routines, scan analysis
Reasoning	`<think>...</think>` tags for chain-of-thought before answering
Recommended	`Q4_0` (97.5% pass, 448 MB)

Evaluation Results — 118-Question Strict Scoring

All quantizations tested with 118 questions across 9 categories (balanced EN/VI). Multi-criteria strict scoring: empty/short check, </think> format, CJK language detection, garbled pattern detection, repetition loops, answer substance (>30 chars), and category-specific validation (scan ≠ greeting, boundary must decline, identity must mention Acnoryx).

All failures manually verified for accuracy.

Quantization	Size	Passed	Pass Rate	Status
Q4_0	448 MB	115/118	97.5%	✅ Best
IQ4_XS	431 MB	114/118	96.6%	✅ Excellent
Q8_0	768 MB	113/118	95.8%	✅ Pass
Q5_K_M	526 MB	112/118	94.9%	✅ Pass
F16	1.5 GB	110/118	93.2%	Baseline
IQ4_NL	449 MB	110/118	93.2%	✅ Pass
Q4_K_M	462 MB	108/118	91.5%	✅ Pass

Category Breakdown — Q4_0 (97.5%, best quant)

Category	Tests	Passed	Pass Rate
Identity (EN/VI)	12	12	100%
Acne Types & Definitions	20	20	100%
Acne Causes & Triggers	10	10	100%
Skincare Ingredients	10	10	100%
Skincare Routines	8	8	100%
Scan Analysis	12	12	100%
Boundary / Refusal	22	19	86%
Format Checks (think tags)	4	4	100%
Out-of-Distribution (OOD)	20	20	100%

Category Breakdown — IQ4_XS (96.6%)

Category	Tests	Passed	Pass Rate
Identity (EN/VI)	12	12	100%
Acne Types & Definitions	20	19	95%
Acne Causes & Triggers	10	10	100%
Skincare Ingredients	10	10	100%
Skincare Routines	8	8	100%
Scan Analysis	12	12	100%
Boundary / Refusal	22	19	86%
Format Checks (think tags)	4	4	100%
Out-of-Distribution (OOD)	20	20	100%

Known Failure Patterns (common across quants)

Category	Issue	Root Cause
Boundary (CJK)	Responds in Japanese/Chinese/Korean instead of declining	0.6B base model lacks language constraint training data
Acne Types	"Mụn thịt có nguy hiểm không?" (skin tag danger)	Confuses skin tags with other conditions
OOD	"Bơi lội có làm mụn nặng hơn không?"	Misreads Vietnamese "bơi lội" (swimming) as "bôi mủ" (applying pus)
Boundary	Math problems (2x + 5 = 11)	Some quants solve instead of declining

Note: CJK boundary failures are the primary weakness. The model responds competently in the query language instead of refusing. This is inherent to the 0.6B base model and would require retraining with explicit language constraint data to fix.

Provided Files

File	Size	Description
`acnoryx-0.6b-f16.gguf`	1.5 GB	Full precision (source model)
`acnoryx-0.6b-q8_0.gguf`	768 MB	Highest quality quantization
`acnoryx-0.6b-q5_k_m.gguf`	526 MB	High quality
`acnoryx-0.6b-q4_k_m.gguf`	462 MB	Good quality
`acnoryx-0.6b-iq4_nl.gguf`	449 MB	Good quality (imatrix)
`acnoryx-0.6b-q4_0.gguf`	448 MB	Best pass rate (97.5%)
`acnoryx-0.6b-iq4_xs.gguf`	431 MB	Smallest release (96.6%)

Usage

llama.cpp

./llama-cli -m acnoryx-0.6b-q4_0.gguf -cnv

llama.cpp (API server)

./llama-server -m acnoryx-0.6b-q4_0.gguf --port 8080

Prompt Format (ChatML)

<|im_start|>system
You are Acnoryx AI, a dermatology assistant...<|im_end|>
<|im_start|>user
What are blackheads?<|im_end|>
<|im_start|>assistant
<think>
...reasoning...
</think>
Blackheads are open comedones...<|im_end|>

Training Details


Data	29,726 samples across 94 cleaned JSONL files
Topics	Identity, acne types/causes/treatments, skincare ingredients/routines, scan analysis, boundary refusals, language constraints
Method	Unsloth SFT with LoRA (r=16, alpha=16)
Targets	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Hyperparameters	3 epochs, batch 2, grad_acc 8, lr 5e-5, warmup 0.08, bf16
Trainable params	~0.74% (6.4M / 859M)
Checkpoint	checkpoint-9795

Limitations

Only supports Vietnamese and English — may respond in other languages if prompted in CJK
Not a medical device — all advice is for reference only
May hallucinate product details, timelines, or diagnoses in edge cases
Skin tag (mụn thịt) knowledge is incomplete for some quants
0.6B parameter count limits complex reasoning ability

Downloads last month: 3

GGUF

Model size

0.8B params

Architecture

qwen3

Hardware compatibility

4-bit

5-bit

8-bit

16-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Acnoryx/Airy-Core-0.6B

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Quantized

(297)

this model

Evaluation results

Q4_0 Pass Rate (118 questions)
self-reported

97.500
IQ4_XS Pass Rate (118 questions)
self-reported

96.600
Q8_0 Pass Rate (118 questions)
self-reported

95.800