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.

Code: Apache-2.0 (LICENSE). Weights: non-commercial research and evaluation only; see each model card's license.

What it does

• noul: probability of yes.
• choice: picked option, probabilities, confidence.
• score: expected level across 2-10 ordered levels, probabilities, confidence.

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. Offer "none" or "unclear" when those can happen.

Get started

Try it

Open the Space.

See Computer use for browser agents and self-hosted screenshots, or try Screen click.

Run ten offline computer-use scenarios against a self-hosted screenshot-enabled blink server, or use the CPU mock to check the harness.

Run it

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

Check curl -s http://localhost:8000/healthz.

For self-hosted latency and throughput, see serving options.

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

blink_mlx.py uses the model repo's blink.py for prompts and answers, and MLX for the forward pass. Text and JSON state only, bf16 weights only, no server. It was checked on Linux, so run --check once on your Mac. See examples/mlx for memory, checks and limits.

Customize

Fine-tune blink on your own decisions, keep it calibrated, and serve the result: the customize guide walks through data, training with PyTorch, on a Mac with MLX or on Foundry, evaluation and RL. The runnable pieces are in examples/train, examples/mlx and examples/rl.

Use it from code

Follow the API docs.

For screenshots on a self-hosted v1.3 or later server, turn on image input.

Use it from agents

See skills/blink/SKILL.md and the agent experience guide.

Models

Code: Apache-2.0. Weights: non-commercial research and evaluation only; see each model card's license.

Links

• Docs
• Agent experience
• Contributing
• Security
• Code: Apache-2.0 (LICENSE); weights: non-commercial research and evaluation only (see each model card's license).