blink
Send text or JSON state with typed questions. Each question gets a probability for every offered option from one forward pass. No generated text.
noul, choice, and score questions.→Train it on your decisionsFollow the customize guide to fine-tune on your data, keep the probabilities honest, and serve the result with PyTorch, on a Mac with MLX, or on Foundry.→Give it to your agentCopy the blink skill, then follow the agent experience guide.→- 1StateA ticket, an email, a document, a row of JSON
- 2Questionschoice · noul · score with your own criteria
- 3Read optionsone forward pass per batch
- 4Probabilitiesnormalised over the options you offered
No text is generated. Every offered option gets a probability.
Large requests may need more than one batch.
Probabilities are over the offered options only, not certified chances of being right.
blink reads each question's prompt in one forward pass and takes the next-token scores of the offered option letters. It never writes text, so it can only pick among the options you list.
Try it in the Space
Open the Space to try blink or use its API tab.
Try Screen click for screenshots, or open Computer use for browser agents and self-hosted setup.
Run the offline computer-use scenarios with a self-hosted blink server; the Computer use scenarios section has the results and videos.
Quickstart
pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
hf download thegovind/blink-4b --revision v1.4 --local-dir blink-4b
python blink-4b/serve.py --model ./blink-4b --port 8000
Run it on a Mac
pip install "mlx-lm>=0.31.3"
curl -O https://raw.githubusercontent.com/thegovind/blink/main/examples/mlx/blink_mlx.py
python blink_mlx.py --model thegovind/blink-4b
python blink_mlx.py --check # compare with the Space's saved runs
Text and JSON state only, bf16 weights only, no server. It was checked on Linux, so run --check once on your Mac. The MLX example covers memory, checks and limits.
Code: Apache-2.0. Weights: non-commercial research and evaluation only; see each model card's license.