blink-4b

Small, fast decisions for routing and checks at volume. Send text or JSON state with choice, noul (yes/no), or score questions. Get a probability for every offered answer, not generated text. Each batch takes one forward pass; large requests can use several batches.

Use choice to route a request, noul for a yes/no check, or score for an ordered rating. The same call can ask several questions about a single state.

Try it: Space · Screen click · Computer use · API · Docs · GitHub · blink-mimo-9b · blink-27b

At a glance

Attribute	Detail
Base model	Qwen/Qwen3.5-4B, text weights only
Weights size	8.4 GB (bf16, 4,205,751,296 parameters)
Revision	v1.4 (code revision; weights identical to v1.0)
License	Weights: non-commercial research and evaluation only (LICENSE.md); code: Apache-2.0. Base-model notice: Apache-2.0 (LICENSE-Qwen).

Quickstart

# 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-4b"
os.environ["BLINK_REVISION"] = "v1.4"
sys.path.insert(0, os.path.dirname(hf_hub_download("thegovind/blink-4b", "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 with TYPESAFE_BASE_URL; text decisions use the same request and response fields as hosted Jev. Requests run one at a time by default; --batch-window-ms 5 enables cross-request batching. The API reference covers limits and errors.

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
# 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-4b
docker build -t blink-4b . && docker run --rm --gpus all -p 127.0.0.1:8000:8000 blink-4b

Higher throughput (opt-in)

serve_vllm.py (added in v1.4) is an opt-in, text-only server with higher throughput. serve.py stays the default.

Earlier paired loopback measurements (one self-hosted replica, 231 public items): c1 p50/p95 was 52/127 ms on vLLM vs 64/175 ms on serve.py (one request at a time); c16 was 34.7 vs 12.2 completed questions/s (16 concurrent requests).

Earlier peak memory was 68.0 GiB including cache and cold start 180 s; neither was retimed, nor were the other quality sets or c16 rerun. A release build re-passed the 231-item JevBench c1 quality check (80/111 hard, 199/231 total, hard ECE 0.069). Later changes touched only request checks, error handling and startup cleanup, not scoring.

Tested versions:

python -m pip install "torch==2.13.0" "transformers==5.17.0" \
  "vllm==0.30.0" "compressed-tensors==0.17.0" \
  "accelerate>=1.1.0" safetensors huggingface_hub
python -m pip install "flash-linear-attention==0.5.2"

hf download thegovind/blink-4b --revision v1.4 --local-dir blink-4b
cd blink-4b
python serve_vllm.py --model . --port 8000 --quantization auto --max-concurrency 32 --max-num-seqs 32 --max-num-batched-tokens 8192 --gpu-memory-utilization 0.85

Setup, quality limits and caveats. No official JevBench score for blink has been published.

Screenshots (opt-in, self-hosted)

Image input is off by default. Start serve.py with --vision-tower Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a 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.

Results

Local development readout	Result
JevBench public-items proxy	76.5; 80/111 hard; hard ECE 0.067
Decision Index 0.2 balanced skill	37.85

No official JevBench score for blink has been published. The JevBench numbers are public-item development proxies, not an official score, rank, or parity claim. Decision Index is a descriptive local run of the official kit, not a leaderboard submission; training overlap affects its scores.

Model details: architecture, training, data

Architecture and readout

Qwen3.5-4B text backbone: 32 decoder layers (24 Gated DeltaNet, 8 full-attention), hidden 2560, and 4,205,751,296 shipped parameters. The training updated 32.5M 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; and 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 option labels. These are option-conditional probabilities, not certified chances of being right.

Training

Supervised fine-tuning on target distributions; no RL or preference optimization. T3 used lr 3e-5 for 96 steps; T4 used lr 4e-5 for 472 steps. The released weights average T3, T4 step 300, and T4 final.

Stage	Question rows	Mix
T3	23,156	11,352 decision worlds · 3,741 teacher-written rows · 2,579 exact-probability worlds · 2,221 public-source · 1,763 base-model anchors · 1,500 program-generated reasoning
T4	42,360	12,000 decision worlds · 7,860 teacher-written rows · 7,000 judge-style · 6,220 exact-probability worlds · 3,500 base-model anchors · 3,000 program-generated reasoning · 2,780 public-source

Data sources and licences

Source	Licence
MMLU auxiliary train, MMLU-Pro, CommonsenseQA, GSM8K	MIT
AQuA-RAT, Amazon ESCI	Apache-2.0
MedMCQA	Apache-2.0 (dataset card)
SuperGPQA	ODC-BY
WANLI, ContractNLI, BANKING77, GPQA	CC BY 4.0
ARC	CC BY-SA 4.0
BoolQ, Dolly-15k	CC BY-SA 3.0
ANLI	CC BY-NC 4.0
SciQ	CC BY-NC 3.0
iSarcasmEval	MIT (upstream repository licence)
VAST, Humicroedit, OpenBookQA	None stated by source
Code-generated worlds and teacher-written documents (Qwen3.8-27B)	See LICENSE.md

Source-repository licences do not settle rights in every underlying text.

Evaluation and limits

The archived Decision Index 0.1 local run scored 52.12; editions are not directly comparable. Public JevBench items were used for development selection, not training. Training included 281 MMLU-Pro test-partition questions, contaminating that Decision Index 0.2 component; semantic and pretraining overlap cannot be ruled out. English-centric, no chatting or explanations; text in state can sway an answer, especially on long policies and date arithmetic.

For blink.py and default serve.py: 255 options per choice, 2–10 score levels, 131,072 input tokens per question, and 512 questions per request; over-limit requests return 422 without truncation. For image placement use --image-layout first|inline (first is the default); for a renamed folder use --model-name blink-4b. Neither flag enables images on its own.

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).