Vega-U: one 14.3M engine with a port for each frozen reader, from the Vega author
publicado (OK do Fabricio 18/10 BRT).
Nandakishor M, who made Vega, released Vega-U on Hugging Face on 17 October 2026 (15:20 BRT) and announced it on X on 18 October (00:16 BRT) together with a new venture, Embedding Space. The blog post and X post say the problem it solves is retraining a decision head for every backbone: one trained engine now works with several frozen language models.
What is in the repo. nandakishorm/vega-U (Apache-2.0 in the Hub metadata) holds a 14.3M-parameter engine of about 57 MB, plus a small linear “port” for each frozen reader: Qwen3.5-0.8B (at the repo root), Qwen3.5-4B, Qwen3.8-27B and gemma-3-1b-it. The code is the vegaml package (0.11 or later; Apache-2.0, 265 stars on 18 October). The X post says the adapter attaches to “0.8B, 4B, 9B, or 27B” models; the repo lists no 9B port, and it does list Gemma 3 1B.
Author numbers (UnverifiedClaim). Held-out accuracy of 0.811 with the 27B reader (calibration error 0.021, AURC 0.053), 0.791 with the 4B, 0.765 with the 0.8B and 0.738 with Gemma 1B. The ordinal score type sits near 0.61. The 27B figure comes from a split capped at 400 records per source. They are not on JevBench, the Decision Index or Artificial Analysis, and no third party has reproduced them. At the time of reading the model had 12 downloads and 5 likes, and the launch post about 170 likes.
Why it matters, and what it does not change. A shared engine, if it holds up outside the author’s tests, would cut the retraining per backbone that adapter-style projects usually need. But Vega’s own independent results from 17 October (JevBench open weights: 4B Intelligence 16.7) are unchanged, and the new numbers are the author’s. The Vega page has a Vega-U section; its scores stay at 4.8 / 5.9 / 2.