blink-27b

The most accurate blink model for hard typed decisions. It also performed best of the three in local browser-agent runs that passed page elements as text, not screenshots; see browser-agent setup. These local pages are not a benchmark.

Send text or JSON state with choice, noul (yes/no), or score questions. Get a probability for each offered answer without generated text. Each batch uses one forward pass; large requests can use several batches.

For browser agents, send the page and candidate elements as JSON state; make operations and click targets choice questions. The probability map lets an agent pick among its proposed actions. This is a text-element workflow, not autonomous web navigation; for pixels, use the optional image path below or MiMo's Screen click demo.

Try it: Space (the smaller models run live; self-host this one) · Screen click · Computer use · API · Docs · GitHub · On a Mac (MLX) · Customize · blink-4b · blink-mimo-9b

Computer use in ten apps

blink-27b was the most reliable here: 199/230 tasks done (87%) zero-shot, with 100/100 truly risky clicks paused for approval. Each step starts with a screenshot marked with numbered boxes. Three typed questions pick a box, check if the task is done, and flag risk. blink sees pixels, the task, and past actions, never the page code or the answer. For screenshots, it borrows the matching Qwen base vision encoder with --vision-tower. These are original apps built for this showcase, not a public benchmark. Watch the runs · Videos and traces · Code.

At a glance

Quickstart

The example asks several questions about the same state. Replace the message with your page, policy, or workflow text and set criteria to actions your application actually supports. The caller decides what to do with the returned probabilities.

# pip install "torch==2.13.0" "transformers==5.17.0" "flash-linear-attention==0.5.2" "accelerate>=1.1.0" safetensors huggingface_hub
import os, sys
from huggingface_hub import hf_hub_download

os.environ["BLINK_MODEL"] = "thegovind/blink-27b"
os.environ["BLINK_REVISION"] = "v1.4"
sys.path.insert(0, os.path.dirname(hf_hub_download("thegovind/blink-27b", "blink.py", revision="v1.4")))
import blink

out = blink.decide(
    "Order #4411 arrived with a cracked screen. I want my money back, not another one.",
    {
        "intent": {
            "type": "choice",
            "instructions": "What does the customer want?",
            "criteria": {"refund": "Money back", "replacement": "A new unit", "info": "Information only"},
        },
        "urgent": {"type": "noul", "instructions": "Does this need a reply today?"},
        "anger": {"type": "score", "instructions": "How upset is the customer?", "criteria": ["calm", "annoyed", "angry"]},
    },
)
print(out["answers"]["intent"]["probabilities"])

Run it as a server

serve.py serves POST /v1/systemone, GET /v1/models, and GET /healthz. Point TypeSafe's server-side Python or JavaScript SDKs at it using TYPESAFE_BASE_URL for text decisions; its request and response fields match hosted Jev. Cross-request batching is optional with --batch-window-ms 5.

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-27b --revision v1.4 --local-dir blink-27b
python blink-27b/serve.py --model ./blink-27b --port 8000
# TypeSafe SDKs: export TYPESAFE_BASE_URL=http://127.0.0.1:8000 TYPESAFE_API_KEY=any

Or use Docker from the downloaded folder:

cd blink-27b
docker build -t blink-27b . && docker run --rm --gpus all -p 127.0.0.1:8000:8000 blink-27b

The vLLM path did not pass blink-27b's quality check: agreement with serve.py on the 5-option web-action set was 489/500, below the 493/500 limit. Use serve.py.

Run it on a Mac (MLX)

blink_mlx.py keeps this repo's blink.py for prompts, option labels and answers, and runs the forward pass with MLX on Apple silicon. Text and JSON state only, bf16 weights only; it is not a server.

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-27b

The Space has no saved runs for this model. Record your own with space/record_replay.py from the GitHub repo, then compare with --check --reference your-run.json. The MLX guide covers memory and limits.

Customize it

Fine-tune blink-27b on your own decisions with a LoRA adapter. Check and split your rows with examples/train/prepare.py, mix in anchor rows so it keeps its general skills, train with the PyTorch trainer on a GPU, locally or in the cloud, or with examples/mlx/train_mlx.py on a Mac (checked against the PyTorch trainer on a small model, not yet timed on a Mac), and compare accuracy and calibration before and after. The customize guide has the steps, and examples/rl covers learning from outcomes and RL.

Screenshots (opt-in, self-hosted)

Image input is off by default. Start serve.py with --vision-tower Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 to attach the matching tower. This is a self-hosted blink extension. TypeSafe's hosted Jev is text-only.

With --vision-tower, self-hosted blink borrows the pinned Qwen base model's vision encoder while its checkpoint stays text-only; blink-mimo-9b uses its own encoder and runs Screen click.

Put a data:image/png;base64,... URI (JPEG and WebP data URIs work too) inside a state string, or pass data URIs in a top-level images list. Image URLs are never fetched. If loading only blink.py via hf_hub_download, also download graft_keys.py from the same repo and revision beside it.

See Computer use to self-host this model with images; its text-only browser-agent use is covered in the API docs.

Results

No official JevBench score for blink has been published. These public-item results are not an official score, rank, or parity claim. Decision Index is a descriptive local run of the official kit, not a leaderboard submission. Training included MMLU-Pro test-partition questions, so that component is contaminated.

Architecture and readout

Qwen3.8-27B text backbone: 64 decoder layers (48 Gated DeltaNet, 16 full-attention), hidden 5120, and 26,895,998,464 shipped parameters. Training merged 116.7M LoRA parameters at rank 16, alpha 32: q_proj, k_proj, v_proj, o_proj; in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj; gate_proj, up_proj, down_proj. The vision encoder and MTP head were removed: 0 vision tensors, 0 MTP tensors.

Readout softmaxes FP32 next-token logits over the offered labels. These option-conditional probabilities are not certified chances of being right.

Training

Supervised fine-tuning on target distributions, no RL or preference optimization. T2 used 72,700 question rows, lr 5e-5, 230 steps; T4 continued with 74,754 rows, lr 3e-5, 856 more steps.

Data sources and licences

Source-repository licences do not settle rights in underlying texts.

Evaluation and limits

• Training overlap. The archived Decision Index 0.1 local run scored 63.44. Its public train splits overlap the training mix. About 3.2k MMLU-Pro test-partition questions in training contaminate that Decision Index 0.2 component. Semantic and pretraining overlap cannot be ruled out.
• Limits. English-centric; no chatting or explanations. Text in state can sway decisions, and date/number reasoning is a weak case. The default server accepts 255 options per choice and 2–10 score levels; over-limit requests return 422 without truncation. --image-layout first|inline controls image placement (first by default); --model-name blink-27b handles renamed folders. Neither flag enables image input.

License

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

See LICENSE.md for weight terms; the Qwen base model is Apache-2.0 (LICENSE-Qwen).