Decision 3.0

Decision 3.0 (d3) answers your own choice, noul and score questions about text, images and video, zero-shot. Every question in a request sees the same images and videos. The family runs on a GPU and returns no spans.

Start the smallest size on an AMD GPU

vllm-sr serve vllm-sr/d3-edge -e --platform rocm

Sizes

Model	Parameters	Answers	Spans	Reads	Device memory
d3-edge	0.595B	choice, noul, score	—	text, image, video	4 GiB
d3-lite	0.853B	choice, noul, score	—	text, image, video	4 GiB
d3-nano	2.21B	choice, noul, score	—	text, image, video	8 GiB
d3-mini	4.54B	choice, noul, score	—	text, image, video	16 GiB
d3-flash	8.39B	choice, noul, score	—	text, image, video	32 GiB
d3	26.1B	choice, noul, score	—	text, image, video	72 GiB
d3-turbo	26.1B	choice, noul, score	—	text, image, video	72 GiB
d3-ultra	26.1B	choice, noul, score	—	text, image, video	72 GiB

• d3-edge to d3 are the plain sizes. Their answers match the packages' own runtime bit for bit (parity record).
• d3-turbo and d3-ultra are 27B state-once models, below.
• d3-edge is a pruned d3-lite. d3-lite, d3-mini, d3-flash and d3-turbo are also the bases of the larger Vela 3.0 sizes.

No spans

d3 answers no span questions. When a d3 size is the decision model, bind pii and hallucination to a span-capable deployment, such as vela3-fast; each model page shows the configuration.

State-once models: d3-turbo and d3-ultra

vllm-sr/d3-turbo reads a request's input once and answers every question from that reading, so asking many questions about one input costs little more than asking one (record):

• Besides choice, noul and score, it answers box questions (where something is in an image, or in a video at the second given by at) and track questions (a box for each moment of a whole video, 1 or 2 per second) on the System One API.
• A question or an option may hold an image of its own (assets).
• Each question reads up to 32,768 tokens by default and up to 262,144 with options.long_context, on one AMD Instinct MI325X.

vllm-sr/d3-ultra also reads a request's input once and answers the same question types as d3-turbo (record):

• A text-only request reads its list of questions in front of its state, so every question sees what else is asked; a request with images or videos reads as d3-turbo's does.
• A one-question request answers at the temperature the package sets for it.
• Each question reads up to 131,072 tokens by default, on one AMD Instinct MI325X.

Quantized packages

A d3 package with FP8, int4 (GPTQ or AWQ) or NVFP4 weights in the compressed-tensors format, and a MODEL_MANIFEST.json, is served the same way, from a local path or from its repository with the commit to load. The published packages of vllm-sr/d3 are vllm-sr/d3-FP8, vllm-sr/d3-W4A16-AWQ, vllm-sr/d3-W4A16-GPTQ, vllm-sr/d3-NVFP4 and vllm-sr/d3-NVFP4-AWQ.

Package	Keeps of the BF16 answers	Device memory (BF16: 52.2 GB)
FP8	99.08%	31.9 GB
GPTQ int4	97.98%	16.4 GB
AWQ int4	97.76%	16.5 GB
NVFP4 AWQ	96.81%	17.2 GB
NVFP4	96.64%	17.2 GB

Quantization explains the formats, the hardware each needs and how to serve one.

Precision and hardware

• d3 runs in BF16 on a GPU, its published precision, and in FP32 on a CPU, where its BF16 answers depend on the CPU's kernels and its FP32 answers do not. FP32 holds twice the bytes of BF16; on a CPU that computes BF16 natively it takes 1.2 to 3.5 times the time per request, and on an AVX2 CPU, which emulates BF16, it is the faster one. --dtype bfloat16 runs BF16 on a CPU, checked against the BF16 reference of its CPU class (Engine arguments).
• It is validated on AMD Instinct MI300X and MI325X; the model pages list the devices whose reference answers are recorded for each revision.
• The ROCm image's PyTorch has no CPU LAPACK, which the Qwen3.5 backbone needs on a CPU: serve d3 on the GPU, or on a CPU from the CPU image.
• On a GPU, d3 batches text rows and image or video rows in separate lanes, and sizes a merged forward to about 0.3 s of device time (Engine arguments).