NVIDIA Nemotron 3.5 Super VL Model Now Runs Locally via Unsloth
NVIDIA's Nemotron 3.5 Super VL, a large-scale long-context foundation model, is now available to run locally using the open-source Unsloth platform. The update also brings improvements to prompt caching and other developer workflow features.
What changed?
NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter model with a context window of up to 1 million tokens, can now be run locally using the open-source Unsloth desktop app. The model is designed for document understanding, computer use, and long-running agentic tasks. It operates using 64 GB (1-bit) or 128 GB (4-bit) of memory. This release brings local support for Muse Spark Open as well as new features in Unsloth: improved prompt caching, support for text-diffusion LLMs, DeepSpeed training, and other performance optimizations for local agent workflows.
Why does it matter to an everyday developer?
Nemotron 3.5 Super VL gives developers access to a large model with extremely long context and improved generation speed—about five times faster than Inkling Small at 100K input tokens—while maintaining similar reasoning abilities. This enables work on complex tasks such as entire document analysis, extended coding jobs, and persistent AI agents without cloud infrastructure or usage fees. Being open source and running locally offers more control, privacy, and the ability to customize workflows. Developers running coding agents like Claude Code or Codex locally can also see less prompt re-processing and quicker response times within Unsloth.
What can the developer do now?
- 1
Install or update Unsloth Desktop
Download the latest Unsloth Desktop release for your operating system from the official repository.
