blink-4b

Send a text or JSON state and typed questions: choice picks from up to 255 options, noul is yes/no, and score takes 2–10 ordered levels. Each question gets probabilities over its offered options from one forward pass, with no generated text. Long or large multi-question requests may use several batches.

Try it: Space demo · blink-4b · blink-27b · blink-mimo-9b · Source code · Docs · API

At a glance

Attribute	Detail
Base model	Qwen/Qwen3.5-4B (text model only; vision encoder and MTP head removed)
Weights size	8.4 GB (bf16, 4,205,751,296 parameters)
Revision	v1.2 (code revision; weights identical to v1.0)
License	Non-commercial research only (LICENSE.md); base model Apache-2.0 (LICENSE-Qwen)

Results

JevBench: public-item development proxies

Model	Public-items proxy	Public hard (111)	Hard ECE	Probability TVD	Official JevBench v1.4
blink-4b	76.5	80/111	0.067	0.226	no official score published
JevK5 v0.2.0	76.1	79/111 (own runtime: 82/111)	0.068	0.220	62.0
Jev 1.13.0	—	—	—	—	63.3

The JevBench numbers are public-item development proxies, not official scores, and claim no rank or parity. Official scoring requires held-out, judge, and sealed items.

How to read the public-item numbers

• On the public hard items, blink-4b scored 80/111 and JevK5 scored 79/111 in the same runtime; JevK5's own runtime reports 82/111. No hard-accuracy advantage is claimed. blink-4b's 95% Wilson interval is 0.631–0.796, before selection effects.
• Speed comes from each row's serial run, applying JevBench's self-hosted adjustment.
• Cost uses JevBench's 4B tariff ($0.03 per million input tokens) times measured tokens per decision; not a production bill.

Decision Index 0.2 (local run)

Decision Index numbers are local runs of the official kit (commit 19ad28e on 2026-09-25), not leaderboard submissions. The 0.2 run is descriptive: known training exposure stays in the scores, with no leaderboard-style penalty, so it isn't ranked. Training included 281 MMLU-Pro test-partition questions, which contaminate the 0.2 MMLU-Pro score. "Without MMLU-Pro" is a sensitivity check, not a clean score.

Balanced skill	Balanced raw	Breadth skill	Without MMLU-Pro
37.85	53.33	36.78	37.41

Breakdown by area and added benchmarks

Area	Number of benchmarks	Skill	Raw
Knowledge & Reasoning	10	26.4	43.1
Language Understanding	10	47.4	62.7
Retrieval & Classification	7	36.8	54.7
Tools & Automation	6	51.6	60.0
Arts & Human Taste	7	27.2	46.1

Benchmarks added in 0.2:

Benchmark	Metric	Requests	Answered	Raw	Skill
PhishNChips phishing decisions	accuracy	2,000	2,000	63.6	27.3
MMLU-Pro	accuracy	12,032	12,032	52.1	46.1
BBH fixed-option tasks	accuracy	5,507	5,507	63.8	47.5
RAGTruth response-level hallucination	F1 on hallucinated class	2,700	2,700	66.5	43.1
HoVer claim verification	accuracy	4,000	4,000	63.1	26.2
When2Call MCQ	accuracy	3,652	3,652	62.8	50.3
New Yorker caption matching	accuracy	528	528	58.9	48.6

• All 151,034 scoreable requests scored across 40 counted benchmarks in five equal areas.
• Shared requests reuse 0.1 predictions; the 30,419 added requests ran with the frozen evaluated soup at temperature 1.0.
• Point estimates only; no significance, calibration, or latency claims. Do not compare with 0.1 due to edition differences.

Training-row text matches in added requests:

Training stage	Rows in the stage	Rows matching added-request text	From MMLU-Pro	From SuperGPQA	Other
T3	23,156	138	131	6	1
T4	42,360	156	148	7	1

Exact normalised strings >= 30 characters shared by training rows and added requests (excluding strings in >= 20 requests). Counts reflect training rows by stage and source, not unique test questions.

Decision Index 0.1 (archived edition)

Local run of the archived 0.1 suite (132,422 requests across 37 benchmarks; 19 panel benchmarks averaged for headline index; comparison rows from 2026-09-22 leaderboard snapshot).

Model	Size class	Decision Index 0.1	Skill	Breadth
blink-4b	4B	52.12	36.04	34.18
Jev 1.13.0	closed	59.51	46.26	44.79
Jevfire	27B	55.74	40.86	39.45
JoshuaSP diffusiongemma (open-jev)	26B-A4B	55.56	40.84	39.19
Decider 35B-A3B	35B-A3B	54.34	39.37	37.99
Kev 9B	9B	50.48	32.96	30.54
Kev 4B	4B	47.43	28.86	25.67

Area	blink-4b	Jev 1.13.0
Knowledge & Reasoning	49.1	68.8
Language Understanding	61.0	62.3
Retrieval & Classification	30.1	37.0
Tools & Automation	70.2	73.6
Arts & Human Judgment	50.2	56.2

Per benchmark scores (19 panel benchmarks, 0.1):

Area	Benchmark	blink-4b	Jev 1.13.0
Knowledge	MMLU	0.749	0.917
Knowledge	GPQA Diamond	0.372	0.783
Knowledge	GSM8K	0.579	0.799
Knowledge	CRUXEval	0.472	0.730
Knowledge	CLadder	0.637	0.726
Knowledge	ChessBench	0.135	0.172
Language	ContractNLI	0.761	0.717
Language	iSarcasmEval	0.452	0.505
Language	VAST	0.617	0.646
Retrieval	BRIGHT	0.172	0.187
Retrieval	Amazon ESCI	0.431	0.552
Tools	BFCL	0.903	0.958
Tools	ToolRet	0.412	0.450
Tools	RouterBench	0.790	0.799
Arts	BPoMP	0.847	0.906
Arts	Humicroedit	0.605	0.619
Arts	POP909-CL	0.034	0.181
Arts	cfcolor	0.581	0.647
Arts	Habermas Machine	0.443	0.459

Use

# 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"] = "v1.2"
sys.path.insert(0, os.path.dirname(hf_hub_download("thegovind/blink-4b", "blink.py", revision="v1.2")))
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 and GET /v1/models, compatible with TypeSafe server-side Python and JavaScript SDKs (TYPESAFE_BASE_URL), JevBench's typesafe adapter, and Decision Index's http engine. GET /healthz reports startup checks. Optional authentication via --api-key or BLINK_API_KEY requires Authorization: Bearer <key> on API routes (returns 401 if missing or invalid; /healthz stays open). Error bodies provide details in error and detail; over-limit requests return 422 without truncation. Requests run one at a time by default, batching questions into single forward passes. Enable cross-request batching with --batch-window-ms 5 (up to --max-batch-requests 16, --max-queued-requests 64). Excess queued requests return 529 with Retry-After (v1.1 returned 503). See the API reference for details.

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 v1.2 --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

Check health: curl -s http://127.0.0.1:8000/healthz.

Run with Docker:

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

To enable cross-request batching:

  hf download thegovind/blink-4b serve.py blink.py --revision v1.2 --local-dir blink-4b
  python blink-4b/serve.py --model ./blink-4b --port 8000 --batch-window-ms 5

Details

Architecture and training progression

	What ships
Backbone	Qwen3.5-4B text model; 32 decoder layers (24 Gated DeltaNet, 8 full-attention), hidden 2560; 4,205,751,296 shipped text parameters
Tuned	32.5M LoRA parameters, merged before averaging
Final weights	Uniform weight average ("soup") of T3, T4 step 300 and T4 final

LoRA targets (rank 16, alpha 32, every language-model layer): full-attention q_proj, k_proj, v_proj, o_proj; Gated DeltaNet in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj; and every MLP's gate_proj, up_proj, down_proj. Token embeddings, all norms and lm_head stayed frozen. The trained adapters were merged into the text weights.

The vision encoder and multi-token-prediction (MTP) head were cut: 0 vision tensors, 0 MTP tensors.

Readout takes next-token logits from offered option labels in lm_head (single tokens A–Z, then two-letter labels), computed in FP32 and softmaxed over offered letters. The rest of the vocabulary is ignored. These are option-conditional model probabilities, not certified chances of being right.

Training used supervised fine-tuning with cross-entropy against target distributions: exact probabilities, teacher-verified probabilities, or one-hot labels, plus base-model KL anchor distributions on teacher-rejected prompts. Choice and yes/no options reshuffled each epoch; score levels maintained order. No RL or preference optimization.

Training progression across steps:

Step	DI 0.1	Public hard	Hard ECE	Notes
Qwen3.5-4B, zero-shot	44.45 DI-S	0.595	0.131	Baseline before decision training.
T1 (106.7k rows; lr 1e-4; 355 steps)	53.15 DI-S	0.559	—	Public train splits and exact-probability items lifted DI-S but hurt hard items; NLI/classification did not transfer to long documents.
T3 (23,156 question rows; lr 3e-5; 96 steps)	—	0.649	0.089	Restarted from base with worlds, exact probabilities, teacher-written docs, ~10% public replay, and 7.6% base anchors.
T4 (42,360 question rows; lr 4e-5; 472 steps)	—	0.712 (step 300); 0.676 (final)	—	Added judge-style items; earlier checkpoint performed better on hard cases.
Soup (T3 + T4 step 300 + T4 final)	52.12 full	0.721	0.067	Averaged three checkpoints for hard accuracy and calibration; shipped.

Negative results: distilling 27B answers into 4B did not help on hard items; a DI-focused 4B gained +0.4 on DI-S; mixing JevK5 weights into the soup did not help; Qwen3.5-9B with the T1 recipe scored 54.5 DI-S.

Training data and data licenses

Training data mix

Stage	Question rows	Mix
T3	23,156	11,352 decision worlds · 3,741 teacher-written question rows · 2,579 exact-probability worlds · 2,221 public-source (~10%) · 1,763 base-model anchors (7.6%) · 1,500 program-generated reasoning
T4	42,360	12,000 decision worlds · 7,860 teacher-written question rows · 7,000 judge-style · 6,220 exact-probability worlds · 3,500 base-model anchors (8.3%) · 3,000 program-generated reasoning · 2,780 public-source

T4 judge-style includes 3,000 GSM8K-train solution checks, 2,500 Dolly-15k routing, and 1,500 program-answer checks. Teacher documents and typed questions were generated by Qwen3.8-27B and kept only when a blind re-solve by the same teacher agreed.

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

These are source-repository licences; they do not settle rights in every underlying text.

Evaluation notes and limits

Evaluation notes

• Held-out and selection: Candidate checkpoints and prompt format were selected on DI-S (3,000 requests). blink-4b was fixed using JevBench development proxies before its full-suite run. It scored 52.29 on the 129,422 requests outside DI-S. Public JevBench items (231 items) and JevK5's 65 hand-written hard items served as development selection sets; none were included in training. A one-time lockbox of 396 held-out authored items from domains unseen in any of the three checkpoints' training data (same generator families, not JevBench's sealed set) scored 0.861 accuracy and 0.026 ECE.
• Training overlap and audit: Public train splits also used by the 0.1 index include ContractNLI, iSarcasmEval, VAST, Amazon ESCI, Humicroedit, and GSM8K train split. ANLI and BANKING77 train splits were also used. Public sources included 281 MMLU-Pro test-partition questions and 2 GPQA extended-set questions. An audit of all question rows against the 0.1 suite and JevBench public items found no content matches (all 36 flags were the BANKING77 template); no public JevBench items or chess positions matched. The 13-word passage check didn't search long-document bodies or option text. Semantic or pretraining overlap can't be ruled out, and private JevBench items weren't available to check.
• Temperature: A split held out from T4 suggested T = 0.82, with negligible gain. But 166 of its 401 items were in T3 training, so it isn't held out from the released average. Temperature 1.0 is retained without fitting.
• Limits: English-centric (Arabic task A scored 0.313, task C pairs 0.645 on 0.1; broader multilingual ability is unestablished); does not chat or explain answers; text in state can sway answers; date arithmetic and long policies are the weakest cases. Limits: 255 options per choice, 2–10 score levels, 131,072 input tokens per question (longest evaluated prompt: 37,906 tokens) and 512 questions per request; over-limit requests get HTTP 422 with the reason, never truncated.

License

Code (blink.py, serve.py, Dockerfile): Apache-2.0. Weights: non-commercial research only; see LICENSE.md. Base model: Apache-2.0 (LICENSE-Qwen).