Small models for real jobs.

Open weights on Hugging Face. Download a model and run it on your own hardware.

heack/Math-Physics-Qwen3.6-10.5B-GGUF

A text model for mathematics and physics, pruned and distilled from Qwen3.6-35B-A3B. It answers in English and Chinese and fits one 16 GB graphics card.

Parameters
10.5 BFrom a 35 B base model
MATH-500
80%32 of 40, levels 4–5; a small fixed subset, Q8
GPQA-Diamond physics
80%24 of 30; a small fixed subset, Q8
Download
6.48 GBQ4_K_M; Q8_0 is 11.17 GB
CC BY-NC-SA 4.0English · ChineseText onlyGGUF · llama.cpp
Run it · Windows PowerShell, NVIDIA GPU
py -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade huggingface_hub
.\.venv\Scripts\hf.exe download heack/Math-Physics-Qwen3.6-10.5B-GGUF g64full_kd-Q4_K_M.gguf examples/chat.py setup-runtime.ps1 --local-dir math-physics
cd math-physics
powershell -NoProfile -ExecutionPolicy Bypass -File .\setup-runtime.ps1
.\runtime\llama-server.exe -m .\g64full_kd-Q4_K_M.gguf -ngl 99 --host 127.0.0.1 --port 18085 -c 8192 -np 1 --jinja --reasoning-format deepseek -fa on -ctk q8_0 -ctv q8_0 -a math-physics-10.5b

heack/JEVision

Give JEVision an image, a question and possible answers. It returns a probability for each answer, for fast visual decisions.

Median latency
356 ms100 local requests, RTX 5060 Ti 16 GB
Accuracy
94.14%2,423 questions; our own test, not an independent benchmark
Download
135 MBAdapter and decision head
Base model
Qwen3.5-4BDownloaded during setup
CC BY-NC-SA 4.0English · ChineseDocker + NVIDIA GPU
Run it · Linux or WSL2
python -m pip install --upgrade huggingface_hub
hf download heack/JEVision --local-dir jevision
cd jevision
docker build -t jevision-local .
docker run --rm -v "$PWD:/release" -w /release jevision-local download_base.py --root /release/data
docker run --rm --gpus all -v "$PWD:/release" -w /release jevision-local example.py --root /release/data --checkpoint /release

heack/HeackMT5-ZhSum100k

A fine-tuned mT5 model that writes short summaries of Chinese text. Trained mainly on Chinese financial news, about one million lines, for 10 epochs.

ROUGE-1
56.46
ROUGE-2
45.81
ROUGE-L
52.98
Base model
mT5
CC BY-NC-SA 4.0ChineseTransformers
Python · transformers
from transformers import MT5ForConditionalGeneration, T5Tokenizer

model = MT5ForConditionalGeneration.from_pretrained("heack/HeackMT5-ZhSum100k")
tokenizer = T5Tokenizer.from_pretrained("heack/HeackMT5-ZhSum100k")

text = "..."  # Chinese text to summarize
inputs = tokenizer.encode("summarize: " + text, return_tensors="pt", max_length=512, truncation=True)
summary_ids = model.generate(inputs, max_length=150, num_beams=4, length_penalty=1.5, no_repeat_ngram_size=2)
print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))

heack/HeackMT5-ZhCleanText1ML

Removes garbled characters and gibberish from Chinese web text while keeping the original content as intact as possible. Long passages in other languages, such as English, are left alone.

Training data
~1M lines
Eval loss
1.706
Base model
mT5
Prompt prefix
filter:
CC BY-NC-SA 4.0ChineseTransformers
Python · transformers
from transformers import MT5ForConditionalGeneration, T5Tokenizer

model = MT5ForConditionalGeneration.from_pretrained("heack/HeackMT5-ZhCleanText1ML")
tokenizer = T5Tokenizer.from_pretrained("heack/HeackMT5-ZhCleanText1ML")

text = "..."  # Chinese text with garbled characters
inputs = tokenizer("filter:" + text, return_tensors="pt")
outputs = model.generate(inputs.input_ids, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Free to learn and build with. Simple to license for business.

  1. Noncommercial use is free under CC BY-NC-SA 4.0, with attribution and share-alike.
  2. Commercial use needs a license: US$1,000 once for a perpetual license to the purchased version of one model. Each model is licensed separately.
  3. Pay with Buy Me a Coffee in USD, and include the model ID, your Hugging Face username and the licensee's name in the payment message.
Covers
Self-hosted commercial use
Not included
Hosting and API usage
Questions
Contact us

Using a model for free? Buy us a coffee.

Every coffee pays for GPU time and keeps the next model open.