Vega
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.8/10
0.8B and 4B in one repo: frozen Qwen3.5 backbone plus a trained engine (57 / 124 MB) and adapters; runs on a T4 or a Mac. Hub metadata now says Apache-2.0, but the repo has no LICENSE file. Own API scripts, no /v1/systemone server. Solo author.
- Capability5.9/10
Choice, boolean and score with images and a 73,728-token context. JevBench open weights (independent): 4B Intelligence 16.7, calibration 58.7, p50 0.19 s; 0.8B Intelligence 3.1. The card's own tables show Vega behind Jev (JevBench 0.662 vs 0.866). Author wins are on hand-picked tasks.
- Adoption2/100
About 724 downloads and 37 likes in eight days; the 15 Oct comparison post drew about 790 likes and 25k views. No integrations or production users found.
Vendor claims
- 4B: soft accuracy 0.725 and calibration error 0.019 on LocalLLaMA/typed-decisions test (0.8B: 0.680 / 0.026); 0.267 s per decision end to end on a T4[Vendor claim — not independently verified]
- 15 Oct post: Vega 0.8B beats Jev on phishing (75.4% vs 61.9%, 800 emails) and news topics (95.5% vs 80.6%), 74 ms vs 162 ms per decision, calibration error 0.06 vs 0.22; the post text says 75 ms on an Apple M5[Vendor claim — not independently verified]
- Vega-U: held-out accuracy 0.811 with the 27B reader (calibration error 0.021, AURC 0.053), 0.791 (4B), 0.765 (0.8B), 0.738 (Gemma 3 1B); one 14.3M engine shared across readers[Vendor claim — not independently verified]
News covering this model
Licence: check before you use it. On 9–10 Oct the repo declared no licence. On 12 Oct its Hugging Face metadata says apache-2.0, but there is still no LICENSE file in the repo and the card text does not state terms. Treat Apache-2.0 as declared in metadata only.
The language model stays frozen; what is trained is an “engine” (14M parameters for 0.8B, 31M for 4B) plus small task adapters. Long states are split into 16 pieces and pooled. The 4B lives under 4b/ in the same repo (vegaml.load("4b")). Moved from our watch list to the catalog on 12 Oct: the engine is trained decision weights.
Independent result and the 15 Oct comparison (17 Oct 2026)
publicado (OK do Fabricio 17/10 BRT).
The post. On 15 Oct the author posted a table “vega vs Jev vs laya” (X) with Vega 0.8B ahead on phishing (75.4% vs Jev 61.9%), news topics (95.5% vs 80.6%), calibration error (0.06 vs 0.22) and latency (74 ms vs 162 ms). The same post says “75 ms in Apple M5 16 GB”. All of it is an UnverifiedClaim. The thread’s replies ask for the link, for cold versus warm latency and which quantization; none were answered in the replies we could read. The post says laya wins on speed (28 ms) and that Jev still leads on agentic tool calling.
What the card says. The phishing (800 emails, item-paired against Jev 1.13.0) and news-topic (ag-news in the S1MB suite) rows are in the card, which labels them “the tasks chosen because Vega leads there”. The 74 ms is a T4 figure for the 0.8B (the 4B is 106 ms), not an M5. The card’s broader tables are less flattering: on its 43-task public menu Vega 4B scores 0.717 accuracy against 0.836 for Jev; on JevBench (231 items) 0.662 against 0.866, and 0.513 against 0.730 on the hard tier. We could not find the post’s calibration figures (0.06 vs 0.22) in the card, whose own JevBench expected calibration error is 0.169 for the 4B and 0.067 for Jev. The post’s “32,000” context for Jev is the default per-question limit; Jev allows 64k per request and 262k with the long-context header.
Independent boards. JevBench open weights (draw v1.6.9) lists both sizes:
| Capability | Intelligence | Calibration | p50 (self-hosted, on-GPU) | Cost / 1,000 decisions | |
|---|---|---|---|---|---|
| Vega 4B | 37.7 | 16.7 (below the board’s gate) | 58.7 | 0.190 s | $0.058 |
| Vega 0.8B | 33.0 | 3.1 (below the gate) | 62.9 | 0.085 s | $0.019 |
| Jev 1.13 (reference) | 77.1 | 63.6 | 90.6 | 0.24 s (API) | $0.032 |
The board’s 0.085 s for the 0.8B is consistent with the author’s 74–85 ms on a GPU; it says nothing about an Apple M5. Vega is not on the Decision Index (only the unrelated vLLM-SR “Decision 2.0 Vega 27B” is). On the independent measure the models are well below Jev, as the card’s own JevBench numbers also show.
Earlier score working (12 Oct 2026, scored on the 4B; superseded above)
- Maturity = 7 × 0.30 + 7 × 0.25 + 3 × 0.25 + 1 × 0.20 = 2.10 + 1.75 + 0.75 + 0.20 = 4.80 → 4.8
- Capability = 9 × 0.30 + 7 × 0.25 + 6 × 0.25 + 5 × 0.20 = 2.70 + 1.75 + 1.50 + 1.00 = 6.95 → 7.0
- Adoption = 4 × 0.35 + 1 × 0.35 + 0 × 0.30 = 1.40 + 0.35 + 0 = 1.75 → 2
Availability 7 because the licence is a metadata tag without a LICENSE file. Latency 6 and accuracy 5: author numbers only.
Score update (17 Oct 2026)
Accuracy, calibration and latency now have independent evidence; types and the other pillars are unchanged.
- Capability 7.0 → 5.9 = types 9 × 0.30 + calibration 4 × 0.25 + latency 7 × 0.25 + accuracy 2 × 0.20 = 2.70 + 1.00 + 1.75 + 0.40 = 5.85 → 5.9.
- Accuracy 5 → 2: JevBench open-weights Intelligence 16.7 against Jev’s 63.6, and the card’s own JevBench table agrees on direction (0.662 vs 0.866). Calibration 7… was 5 in the 12 Oct working and is now 4: board calibration 58.7 against 90.6, and the card’s own ECE (0.169 on JevBench) contradicts the 0.019 claim from the typed-decisions test. Latency 6 → 7: the board measured 0.190 s p50 on-GPU (4B), no p99 (precedent for board-measured self-hosted p50).
- Maturity 4.8 unchanged. Adoption 2 unchanged (engagement 6 × 0.35 + ecosystem 1 × 0.35 + production 0 = 2.45 → 2).
Vega-U: one engine for several readers (18 Oct 2026)
publicado (OK do Fabricio 18/10 BRT).
On 17 Oct the author released nandakishorm/vega-U (Apache-2.0 in the Hub metadata; 12 downloads and 5 likes on 18 Oct) and announced it on X on 18 Oct together with a new venture, Embedding Space. It is a 14.3M-parameter engine (about 57 MB) shared across frozen readers, each with a small linear port: Qwen3.5-0.8B (at the repo root), Qwen3.5-4B, Qwen3.8-27B and gemma-3-1b-it. It runs through vegaml 0.11 or later (Apache-2.0, 265 stars).
Author numbers (UnverifiedClaim). Held-out accuracy 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 split is capped at 400 records per source. The X post says the adapter fits “0.8B, 4B, 9B, or 27B” models; we found no 9B port in the repo.
Scores unchanged (4.8 / 5.9 / 2). Vega-U is on no board and no third party has run it. The independent result on this page (JevBench open weights, 4B Intelligence 16.7) still stands. See news.
Update, 19 Oct 2026 (publicado (OK do Fabricio 19/10 BRT)): Vega-U on JevBench. The open-weights board (draw v1.6.29) now lists Vega-U with the Qwen3.5-0.8B reader: capability 36.5, Intelligence 5.5, calibration 67.5, p50 0.695 s, $0.012 per 1,000 decisions. It is outside the board’s latency cap (2.5× Jev) and gets no official rank. Next to the author’s held-out 0.765 for the same reader, it repeats the gap we saw with the 4B. Scores unchanged (4.8 / 5.9 / 2): this page is scored on the 4B, whose board result already set accuracy at 2.