Decision 1.0

Decision 1.0 is the first family of decision models: three 0.6B encoders (Kai, Lex and Route, on the Vela encoder backbone) and four decoders on Qwen3.5 (Eos-0.8B, Sol-2B, Nox-4B and Lux-9B). They answer your own choice, noul and score questions about text, zero-shot.

Start an encoder on a CPU

vllm-sr serve vllm-sr/Decision-1.0-Kai-0.6B --engine --platform cpu

Sizes

Model	Parameters	Answers	Spans	Reads	Device memory
Decision-1.0-Kai-0.6B	0.572B	choice, noul, score	—	text	4 GiB
Decision-1.0-Lex-0.6B	0.572B	choice, noul, score	—	text	4 GiB
Decision-1.0-Route-0.6B	0.572B	choice, noul, score	—	text	4 GiB
Decision-1.0-Eos-0.8B	0.753B	choice, noul, score	—	text	4 GiB
Decision-1.0-Sol-2B	1.88B	choice, noul, score	—	text	8 GiB
Decision-1.0-Nox-4B	4.21B	choice, noul, score	—	text	16 GiB
Decision-1.0-Lux-9B	7.94B	choice, noul, score	—	text	24 GiB

tip

For new configurations, prefer Decision 2.0 or Decision 3.0. Decision 1.0 stays built in, so configurations that name it keep working.

How it runs

• Byte-identical to the released runtime. With each request alone, the family answers byte-identically to each package's bundled runtime, for all seven packages, on the Router's ROCm image (parity record).
• Precision: the encoders run in FP32 everywhere; the decoders in BF16 on a GPU and FP32 on a CPU.
• Faster encoders on x86 CPUs: Kai, Lex and Route run merged forwards on pre-packed FP32 copies of their linear layers with oneDNN, 1.5–1.7× faster with no decision changed.
• Shared context is off for the decoders: their records show decisions changing, so --enable-shared-context is an explicit choice (Engine arguments).
• No spans. Bind pii and hallucination to a span-capable deployment when a Decision 1.0 model is the decision model.