Skip to content

Releases

v0.1.0-rc1: edge subset for Jetson Thor

Pre-release

Choose a tag to compare

@TieJean TieJean released this 09 Oct 18:19

Edge subset for timing vision-language models and measuring their memory on NVIDIA Jetson Thor.

Built from commit 1e1f880, from full runs on 1× NVIDIA RTX PRO 6000 Blackwell Server Edition.
Asset edge-subset.tar.gz: SHA-256 4e18cc2b44ef8ca80a6951f79aaaf6cdfbbec860e00b7dcacdfe1370a0d2baca

53 requests per model, covering RoboSpatial-Home, InstructPart, RoboBench, EPIC-Bench, OVDEval and
CaP-X; each may use 16,384 tokens.

Model Weights Memory reserved Estimated time on Thor (full-run answer lengths)
cosmos-reason2-2b-nvfp4_nothink vrfai/Cosmos-Reason2-2B-NVFP4 9.5 GiB 2 min
cosmos-reason2-8b-nvfp4_nothink vrfai/Cosmos-Reason2-8B-NVFP4 17.8 GiB 3 min
qwen3.5-0.8b-nvfp4_nothink AxionML/Qwen3.5-0.8B-NVFP4 4.4 GiB 2 min
qwen3.5-4b-nvfp4_nothink AxionML/Qwen3.5-4B-NVFP4 9.3 GiB 3 min
qwen3.5-9b-nvfp4_nothink AxionML/Qwen3.5-9B-NVFP4 16.2 GiB 5 min
qwen3.8-27b-nvfp4_nothink nvidia/Qwen3.8-27B-NVFP4 35.6 GiB 8 min

How to run: README.md, section "Speed and memory on an edge device (NVIDIA Jetson Thor)".