Decision 2.0's Vega card prints 56.5 on its own Decision Index run

Xunzhuo Liu posted Decision 2.0 on October 3, 2026, at 01:25 UTC. The vLLM account had posted the same release two minutes earlier. The models belong to vLLM Semantic Router. A call takes text or JSON, plus questions whose answers were listed in advance, and returns a probability for every option. The cards say the questions in one request are scored in one forward pass, and that the model does not write a reply. The license line on each card is Apache-2.0. Other calls of this shape are on decision models like Jev.

The six cards were opened October 5. Each prints JevArena, a human-labelled transfer score, and a Jev Decision Index cell. The index footnote says the Decision 2.0 cells are an independent reproduction with the official 0.2.1 kit on the released weights. Other models in those tables are a public board snapshot from September 28, 2026. The footnote says training data was audited at row level against all Index test items. No table on these cards includes Jev. The September 28 board’s Jev cell, 57.91, stays on the Decision Index page.

Model	Parameters	Context	JevArena	Transfer	Index	Median
Kai	0.60B	8,192	48.6	45.9	16.3	4.9 ms
Eos	0.75B	16,384	53.9	50.3	20.1	6.0 ms
Sol	1.88B	16,384	52.1	51.3	29.5	7.2 ms
Nox	4.21B	16,384	63.6	52.3	43.8	12.9 ms
Lux	7.94B	16,384	68.1	56.2	46.3	18.4 ms
Vega	29.37B	32,768	74.0	58.7	56.5	71.4 ms

The names say 0.6B, 0.8B, 2B, 4B, 9B, and 27B. The parameter column is the figure each card prints. Vega is an adapter on Qwen3.8-27B, and its install line adds peft. Kai is a fine-tune of Qwen3-0.6B-Base. Eos, Sol, and Lux are fine-tunes of the Decision 1.0 checkpoint with the same name. Nox is a fine-tune of Qwen3.5-4B-Base. The median is one question on a single GPU. The cards do not name that GPU.

JevArena’s footnote says every model answers the same frozen prompts, and that a missing or invalid answer counts as an error. It does not say how many prompts that is. Transfer is the median macro-F1 over 15 human-labelled tasks, times 100. That footnote is on all six cards.

Vega’s highlight calls 74.0 statistically level with AutoJev-27B at 72.1, and it does not print the interval. On transfer, Vega, AutoJev-27B, and Eikos-27B each print 58.7. Jebadiah-27B prints 57.8. Nox’s highlight calls 63.6 statistically level with Decider 4B at 61.9. On transfer that neighbor is ahead, 55.5 to 52.3. Sol’s JevArena cell is 52.1, below Eos at 53.9. The cards still call each of those two first inside its own size class.

Against the Decision 1.0 checkpoint of the same name, the index cells move from 6.5 to 16.3 for Kai, 18.4 to 20.1 for Eos, 25.3 to 29.5 for Sol, 34.4 to 43.8 for Nox, and 43.5 to 46.3 for Lux. The JevArena cells on that pairing move from 35.9 to 48.6, 42.5 to 53.9, 45.8 to 52.1, 56.5 to 63.6, and 65.8 to 68.1. Vega has no Decision 1.0 row. Its highlight says the index cell is 11.2 points above Lux. The printed cells, 56.5 and 46.3, differ by 10.2.

Loading on the cards is AutoModel.from_pretrained with trust_remote_code=True, and the install line asks for transformers>=5.17. The method is system_one, with choice, noul, and score questions in one dict. Kai’s decision_config.json prints max_options 255. That file does not print a question cap. The vLLM post says one request can hold 64 questions. The cards do not print 64.

Werdykt v1 already lists Decision-2.0-Nox-4B at 71.64%, p50 863.8 ms, on an H100 NVL. That export is a different test from the 63.6 and 43.8 on Nox’s card. We did not run these weights.

On October 7, 2026 the same account posted Vela 2.0. The 0.8B, 4B, and 9B cards say they are derived from Eos, Nox, and Lux. The 9B card prints 41.63 on Decision Index 0.2.1 with Noul calibration on. The write-up is Vela 2.0.