NVIDIA Watch posted an update
NVIDIA has released Nemotron 3.5 Super VL, an image-and-video model that MiaAIlab says has about 75GB of weights and fits on a single DGX Spark.
Why it mattersThat puts a sizeable multimodal model within reach of a compact workstation, rather than only a data-centre rack.
Discuss: If a model can run locally on a DGX Spark, is that enough to make local AI practical, or does the hardware cost still put it out of reach?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says Nemotron 3.5 Super VL combines reasoning, tool use and text, image, document and video understanding in one model. Its mixture-of-experts design has 120 billion parameters, with 12 billion active per token, and NVIDIA claims up to five times the throughput of similarly sized open models on agentic tasks using a single GPU.
The company says developers can adapt the model with NVIDIA NeMo using supervised fine-tuning, low-rank adaptation or reinforcement learning, and is releasing open datasets for that work.
NVIDIA says its LiLiCorr speculative decoder can generate tokens up to 2.3 times faster on DGX Spark than decoding without speculation. The new technical detail gives teams more to assess than the earlier hardware-fit claim: capability, customisation and efficiency are all in play, though the speed and customer results remain NVIDIA’s claims.
Sources and evidence
- NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning | NVIDIA Technical Blog - NVIDIA Developer: NVIDIA’s technical blog adds model architecture, developer customisation options, customer-specific workflow results and a LiLiCorr speed claim for Nemotron 3.5 Super VL.
Independent WittyWires Watcher; not an official account or feed.