ENPT

Blink v0.3 (26B-A4B)

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.9/10

    Open weights under Apache-2.0 in NVFP4 (Blackwell) and FP8 (Hopper/Ada), with a detailed card, a live demo Space, vLLM serving instructions, protocol docs and a system prompt; also bundled in Decision Models (s1-fast, s1-vision). No SLA or hosted endpoint by Pixilab itself; named integrations limited to Nemini.

  • Capability7.6/10

    Official boards (22 Oct): Decision Index 0.4.1 Full 55.92 (#63); JevBench Intelligence 49.2, calibration 90.7, p50 0.051 s. One calibrated probability per option in a single forward pass. No published p50/p99 in ms.

  • Adoption12/100

    ~1,005 downloads and 5 likes on the NVFP4 repo in the month it launched, plus a listed Decision Models host and its own demo Space. Tied with Surogate Rune v3 on the board, ahead of Decider chat by 0.15.

Vendor claims

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

Read this first. Blink is an open decision model that returns one probability per option letter, so it is a choice model even though yes/no questions are encoded by leaving the first option as “No” and the second as “Yes”. All numbers above are the authors’ (UnverifiedClaim).

What it is

Blink v0.3 is a 26B-A4B Mixture of Experts decision model from Pixilab, fine-tuned from google/gemma-4-26B-A4B-it, Apache-2.0. It reads state, question and options and answers in one generated token, taking the softmax over option letters. The card documents the recommended temperature (1.3) and a protocol (“surogate decisions v1”) for interoperability with any vLLM server implementing it. NVFP4 and FP8 builds exist, and the model is available through the Decision Models API under the names s1-fast and s1-vision.

Author results (UnverifiedClaim)

  • Decision Index 0.2.1: 57.48 overall, with Knowledge 42.8, Language 64.3, Retrieval 63.0, Tools 70.0 and Arts 43.8; roughly second among the public entries listed alongside Surogate Rune v3 (57.44).
  • Calibration: 0.739 confidence against 0.735 accuracy, ECE 0.016 over 216,942 scored decisions.
  • Memory: 17.5 GB on disk, 4B active parameters, served on one RTX PRO 5000 with 32k context.
  • Movement since v0.2: iSarcasmEval 58.9 (was 41.8), RAGTruth 59.1 (was 48.3), ForecastBench 28.0 (was 17.4), ContractNLI 66.7 (was 61.1).

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board Row Result
Decision Index 0.3.2 (16 Oct 19:00 UTC) Blink v0.3 26B-A4B (NVFP4) Full 57.76 (#32, tie group 8); public 58.21 (#32); same-skills 57.21; new-domains 50.08
JevBench open weights (live 17 Oct) Blink v0.3 26B-A4B Capability 69.9; Intelligence 49.2; calibration 90.7; p50 0.051 s; $0.048/1k
JevBench open weights (live 17 Oct) blink-4b Capability 61.6; Intelligence 35.3; calibration 88.0; p50 0.057 s; $0.040/1k

Board-status statements on this page to recheck against the table (may be superseded):

  • “Decision Index 0.2.1: 57.48 overall, with Knowledge 42.8, Language 64.3, Retrieval 63.0, Tools 70.0 and Arts 43.8; roughly second among the public ent…”

Score update (audit, 17 Oct 2026). Capability 7.6 → 7.6 (unchanged) = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 5 × 0.25 + 8 × 0.20 = 7.60 → 7.6. Types, calibration and latency come from the earlier working (10 / 7 / 5); only accuracy is re-derived from the board (8; was 8): Decision Index Full 57.76 ÷ 60.11 × 8 = 7.7, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (PixilabAI/Blink-v0.3-26B-A4B-NVFP4): 1,065 downloads, 5 likes, licence tag apache-2.0, last modified 2026-10-06.

Board audit (20 Oct 2026, DI 0.4)

publicado (OK do Fabricio 20/10 BRT).

Decision Index 0.4 (20 Oct; Full now from the private parts only) lists Blink v0.3 26B-A4B (NVFP4) at Full 55.90 (#56); it was 57.76 on the archived 0.3.2 edition. Accuracy rule, reference Jev 1.13 on DI 0.4 Full (61.26): 8 × 55.90 ÷ 61.26 = 7.30 → 7 (was 8). Capability = 10 × 0.30 + 7 × 0.25 + 5 × 0.25 + 7 × 0.20 = 7.40 → 7.4 (was 7.6). Other sub-scores, maturity and adoption unchanged.

Board audit (22 Oct 2026, DI 0.4.1)

publicado (OK do Fabricio 22/10 BRT).

Decision Index 0.4.1 (22 Oct; reference Jev 1.13 Full 59.22, was 61.26 on 0.4) lists this page’s row at Full 55.92 (#63). Accuracy rule: 8 × 55.92 ÷ 59.22 = 7.55 → 8 (was 7). Capability = 10 × 0.30 + 7 × 0.25 + 5 × 0.25 + 8 × 0.20 = 7.60 → 7.6 (was 7.4). Maturity and adoption unchanged.