Published
Updated
Blink v0.3 prints Decision Index 0.2.1 at 57.48
This topic was created 19 days ago, and the information it contains may have evolved or changed since then.
Pixilab's Blink v0.3 card prints a self-run of Decision Index 0.2.1 at 57.48, against the board's Jev cell of 57.91. The card's license field is Apache-2.0. The base is Gemma 4 26B-A4B. We did not run the weights.
Pixilab’s Blink v0.3 card was last modified on October 6, 2026. The repository was created on October 2. Opened October 12, the card says the checkpoint reads a state, a question, and a list of options, and returns a probability for every option from one forward pass and one generated token. The maker line is Pixilab and Nemini. An FP8 build sits in a sibling repository. This page is the NVFP4 card.
The score is a self-run of Decision Index 0.2.1. The card says each copied row and Blink’s own row used 150,759 requests, all answered, and that Blink’s row is not a board submission. The same harness reproduces the board’s Decider 35B-A3B NVFP4 entry at 46.93 against its published 47.11.
| Model | Index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---|---|---|---|---|---|
| Jev (hosted) | 57.91 | 51.4 | 62.0 | 55.4 | 75.1 | 37.7 |
| Blink v0.3 | 57.48 | 42.8 | 64.3 | 63.0 | 70.0 | 43.8 |
| Surogate Rune 26B-A4B v3 | 57.44 | 43.4 | 63.1 | 63.5 | 71.2 | 41.9 |
| Decider chat, Gemma-4-31B | 57.33 | 44.3 | 60.4 | 63.1 | 75.6 | 38.3 |
| AutoJev-27B | 56.40 | 40.9 | 63.5 | 54.9 | 79.4 | 39.4 |
| Blink v0.2 | 55.97 | 42.3 | 60.4 | 63.0 | 69.3 | 41.4 |
| Blink v0.1 | 54.90 | 40.9 | 60.0 | 62.5 | 66.3 | 42.0 |
The card calls 57.48 second among the public entries and level with Rune at 57.44. The gap to the board’s Jev cell is 0.43. The Decision Index kit treats a gap of 0.25 as a tie, so 0.43 sits outside that band. The gap to Rune is 0.04. Edition 0.3’s Jev reference of 60.1, and d3’s 64.7 on edition 0.3.1, are later boards.
Against v0.2, the card prints Language 64.3 from 60.4 and Arts 43.8 from 41.4. The largest moves it names are iSarcasmEval 58.9 from 41.8, RAGTruth 59.1 from 48.3, ForecastBench 28.0 from 17.4, and ContractNLI 66.7 from 61.1. BFCL tool selection is 94.6 from 93.5. When2Call fell to 56.7 from 60.4, against Rune v3 at 68.0. GPQA Diamond is 27.2 from 29.9. HoVer is 58.7 from 60.7. Tools overall is 70.0. Those rows are the card’s own suite notes.
Calibration on the card uses temperature 1.3. Confidence is 0.739 against accuracy 0.735, and the calibration error is 0.016 over 216,942 scored decisions. The card says temperature changes how sure the answer claims to be, and does not change which option wins. v0.2 wanted 1.9. At 1.0 the card says v0.3 is overconfident.
The protocol on the card is one question per prompt, thinking off, and a softmax over the option letters at the first generated position. Yes/no puts No first and Yes second. Past 26 options the codes become two letters. The card says a third of the answers never close the thought if thinking is left on.
The title line is 26B-A4B. The body says 4B active parameters, 17.5 GB on disk, a 32k context, and one RTX PRO 5000. The Hub listing opened the same day reports a safetensors total of 14,386,941,232 parameters, 2,967,950,926 in BF16 and 11,418,992,640 in U8, and the page chrome says 14B. Stored bytes on that listing are 18,820,655,368. The license field is Apache-2.0. The base card for gemma-4-26B-A4B-it also tags Apache-2.0 and links the Gemma 4 license page.
Fyx posted on October 12, 2026, at 00:03 UTC, that Blink plays Pokemon, Mario, Snake, and Tetris, and called the weights Apache 2. An earlier clip the same account posted on October 11 shows Mario. Both attachments are video. We did not transcribe a score. We did not run the weights.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
The card PixilabAI/Blink-v0.3-26B-A4B-NVFP4 was created October 2, 2026 and last modified October 6, 2026. Opened October 12. It prints Decision Index 0.2.1 at 57.48 on the lab's own run of the suite, 150,759 requests, all answered. The Jev cell on that table is 57.91. The card says the run is not a board submission. The same harness reproduces Decider 35B-A3B NVFP4 at 46.93 against a published 47.11.
The license field on the card is Apache-2.0. The base is google/gemma-4-26B-A4B-it, whose own card also tags Apache-2.0 and links the Gemma 4 license page. The card text says 4B active parameters, 17.5 GB on disk, 32k context, one RTX PRO 5000. The Hub listing's safetensors total is 14,386,941,232 parameters. Fyx posted game clips on October 11 and October 12. We did not transcribe a score from the video, and we did not run the weights.
Compare
57.48 and 57.91 are Decision Index 0.2.1. The gap is 0.43. The public kit treats 0.25 as a tie, so this gap sits outside that band. Blink v0.3 against Surogate Rune v3 is 57.48 against 57.44. Edition 0.3's Jev reference of 60.1, d3's 64.7, and Vela's 41.63 are other prints.
When2Call 56.7 is the card's note against v0.2 at 60.4 and Rune at 68.0. Clef's When2Call row of 72.37, and Jev's 80.97 on that blog, are Cloudflare's table. RAGTruth 59.1 against v0.2 at 48.3 is this card's movement. Clef's RAGTruth F1 numbers are a different test.
Terms
- Blink v0.3
- Pixilab's decision checkpoint on Gemma 4 26B-A4B. The card's license field is Apache-2.0. Its self-run of Decision Index 0.2.1 is 57.48, against the board's Jev cell of 57.91.
- Blink readout
- The card says one question per prompt, thinking off, and the answer is a softmax over option letters at the first generated token. The shipped temperature is 1.3.
- Blink parameter counts
- The card text says 4B active parameters and 17.5 GB on disk. The Hub listing's safetensors total is 14,386,941,232 parameters, and the page chrome says 14B.