vLLM Semantic Router releases Decision 2.0, open decision models from 0.6B to 27B
The vLLM Semantic Router team announced Decision 2.0 on the evening of 2 October 2026 (Brasília time). The launch post had about 740 likes, 39,000 views and 680 bookmarks by the next morning. The family is the successor to Decision 1.0 Lux and Kai, with six sizes: Kai-0.6B, Eos-0.8B, Sol-2B, Nox-4B, Lux-9B and a new top model, Vega-27B (an adapter on Qwen3.8-27B). All are Apache-2.0, use the same system_one() call as 1.0, and return a probability for every option without generating text.
The numbers (UnverifiedClaim). From the model cards:
| Model | Parameters | JevArena | Human-labelled transfer | Decision Index (1.0 → 2.0) | Median latency |
|---|---|---|---|---|---|
| Vega-27B (new) | 29.37B | 74.0 (AutoJev-27B 72.1) | 58.7 | 56.5 | 71.4 ms |
| Lux-9B | 7.94B | 68.1 | 56.2 | 43.5 → 46.3 | 18.4 ms |
| Nox-4B | 4.21B | 63.6 (Decider 4B 61.9) | 52.3 (Decider 4B 55.5) | 34.4 → 43.8 | 12.9 ms |
| Sol-2B | 1.88B | 52.1 | 51.3 | 25.3 → 29.5 | 7.2 ms |
| Eos-0.8B | 0.75B | 53.9 | 50.3 | 18.4 → 20.1 | 6.0 ms |
| Kai-0.6B | 0.60B | 48.6 | 45.9 | 6.5 → 16.3 | 4.9 ms |
The Decision 2.0 index scores are labelled “independent reproduction with the official 0.2.1 kit”, but the team ran them itself; the comparison rows are the public board snapshot from 28 September. On those terms Vega would sit just under Jev (57.91). The cards do not quite agree with each other: Vega claims “+11.2 over Lux 2.0”, while the two published scores differ by 10.2. JevArena, which each card leads “for its size”, is the team’s own comparison of same-size open models, and Nox-4B trails Decider 4B on the human-labelled transfer column. Latencies are per single-question request on one unnamed GPU.
What changes here. The family has its own catalog page, Decision 2.0 (4.9 maturity / 6.7 capability / 7 adoption), scored on the same terms as other self-run releases: accuracy stays at 6 until 2.0 is on the public index. The Decision 1.0 Lux and Kai pages keep their official-index scores and now point to their successors. Each 2.0 model had about 70–130 downloads on 3 October.
Update (4 Oct). The launch claims reached a wider audience through a summary by @TeksEdge (David Hendrickson, 3 Oct; about 108 likes and 5.8k views) and the team’s LinkedIn post. Both add claims that are not in the cards (UnverifiedClaim):
- “#1 at its size” on the Decision Index at 0.6B to 4B. This is self-run against the 28 Sep board, and the 2B and 4B leads are under one point.
- “The 27B is #3 overall”. That holds only among open systems; counting Jev, Vega would be #4.
- “64 questions in 63 ms” and “20× faster than the Jev API”. Neither names a model size or GPU.
Details are on the Decision 2.0 page. No score change.