Ollama v0.40.5 runs NVIDIA Nemotron 3.5 Bolt, a 74B vision model sized for one 24 GB GPU
Ollama now runs NVIDIA Nemotron 3.5 Bolt, a 74 billion parameter vision model tuned to fit one 24 gigabyte NVIDIA card. Mac owners have to wait, because Apple silicon isn't supported yet.
Nemotron 3.5 Super reads images and up to a million tokens
$ ollama run nemotron-3.5-super
Bolt is distilled from Nemotron 3.5 Super, which has 120 billion parameters. Super reads documents and images, calls tools, and takes up to a million tokens of context, but the download is 80 to 90 gigabytes.
On NVIDIA, the MLX runner processes Nemotron prompts about twice as fast
On NVIDIA cards, Ollama's MLX runner now has CUDA kernels for Nemotron's Mamba layers, where it used to fall back to slower generic code. Prompt processing roughly doubles, and generation gets up to a third faster.
Clef and Clef Flash decision models run on Ollama's cloud
Clef and Clef Flash decision models run on Ollama’s cloud
clef:cloud clef-flash:cloud
The Clef and Clef Flash decision models now run on Ollama's cloud. Add the cloud tag to either name and the model runs on ollama.com, so your machine skips the download.
Fixes for Responses tool calls and cloud errors
- Responses API conversations with custom tool calls continue without an error
- Cloud errors name the problem: sign-in, payment or ChatGPT gateway access
- Cloud requests containing <, > and & keep their real size
Continuing an OpenAI Responses conversation that used custom tool calls no longer fails. Cloud errors now say whether sign-in or payment is the problem. The rest is in Ollama's release notes on GitHub.
