Vela 2.0

Vela 2.0 is the previous family of Router models. Besides choice, noul and score, it answers questions that pick several labels (set) or mark spans of text (span), and the Router routes on both. Every size has calibrated thresholds for the built-in questions.

Restore the previous default on a CPU

vllm-sr serve vllm-sr/Vela-2.0-0.3B --platform cpu

Sizes

Model	Parameters	Answers	Spans	Reads	Device memory
Vela-2.0-0.3B	0.309B	choice, noul, score, set, span	router	text	4 GiB
Vela-2.0-0.8B	0.756B	choice, noul, score, set, span	router, broad	text	4 GiB
Vela-2.0-4B	4.21B	choice, noul, score, set, span	router, broad	text	16 GiB
Vela-2.0-9B	7.95B	choice, noul, score, set, span	router, broad	text	24 GiB

• On a CPU, run the 0.3B, a ModernBERT encoder that reads up to 8,192 tokens. The 0.8B is a decoder: on a CPU a request takes seconds.
• On a GPU, the larger sizes read inputs of up to 16,384 tokens: the 0.8B costs the least of them, and the 4B and 9B are the most accurate.
• Span heads: every size has the router span head, which answers the pii and hallucination signals with its own ready-made questions; the decoders add the broad head.

Choose a model has each size's accuracy, latency and calibrated thresholds.

Vela 2.0 0.3B, the previous default

Naming vllm-sr/Vela-2.0-0.3B as the primary deployment's artifact restores it, and the calibrated thresholds follow it (jailbreak 0.75, domain 0.28, PII 0.01, fact check 0.93, user feedback 0.37).

When it became the default, the maintainers chose the 0.3B although it missed two goals they had set for it (#4639): level or better accuracy on every signal, and level or better latency on a CPU. Measured through the Router on the router signal suite (A/B record):

• Ahead: prompt guard (held-out AUC +0.026; on the E2E attack fixtures it blocks all six attacks, Vela 1.0 Guard five) and safety (+0.052 held-out, and ahead on every set). The signals share one model.
• Level: PII and hallucination on held-out and fresh files.
• Behind, most: modality (held-out AUC −0.180; the 0.3B misses most requests that ask for a new image) and user feedback (accuracy −0.038 held-out, −0.178 fresh).
• Behind: domain (accuracy −0.037 held-out, −0.088 fresh) and fact check (held-out AUC −0.101).
• CPU time: each request carries the questions, their options and the 17 PII labels (at least 560 tokens) through the 307M-parameter model, where each Vela 1.0 model reads only the request. On 12 CPU cores, for the five request signals of the latency record, the median request takes about 4.9 times as long:

Router on 12 CPU cores	p50	p95	Requests per second	At concurrency 16
Vela 1.0 specialists (since removed)	16 ms	58 ms	38.9	51.8
Vela 2.0 0.3B (the previous default)	79 ms	100 ms	11.9	12.8

#4668 works on the CPU latency. On a GPU, the 0.3B answers the same questions in about 7 ms at the median on one AMD Instinct MI325X (measurements).

note

The Vela 1.0 specialist classifiers answer no System One question, so a binding cannot name them; every application signal asks a decision model.