Ollama 0.40.5 shipped on October 14, 2026, and the headline is NVIDIA’s Nemotron 3.5 family landing for local (and cloud-adjacent) runs. Super is the big sibling; Bolt is the one that fits a 24 GB NVIDIA GPU. If you self-host agents, this is a bookmark-worthy tag.
What Super and Bolt actually are
Nemotron 3.5 Super is a vision-language model aimed at coding, reasoning, tool use, and document understanding. The release notes describe a 120B-parameter hybrid Mamba-2/Transformer MoE backbone with about 12B active per token, a C-RADIO vision encoder, tool calling, an on/off thinking mode, and context up to 1M tokens. Run it with ollama run nemotron-3.5-super.
Nemotron 3.5 Bolt distills that down to about 74B parameters with 9B active, tuned to fit ~24 GB GPUs for everyday agentic work, including speculative decoding via an included drafter. That is the practical “can I actually host this” line for a lot of home and lab boxes — less romance about trillion-parameter clouds, more honest VRAM math.
Other fixes worth knowing
0.40.5 also puts decision models on Ollama cloud (clef:cloud / clef-flash:cloud), fixes continuing an OpenAI-compatible Responses conversation when custom tools were called, and clarifies cloud error messages (sign-in, payment, ChatGPT gateway). llama.cpp and MLX got updates for Nemotron support.
The Responses continuity fix is the quiet win for tool-using agents: fewer mid-thread failures when a custom tool call is already in the conversation history. Pull the tag, read the notes, then decide whether Super’s context window or Bolt’s VRAM fit matches your box. I care less about launch-day hype and more about whether the model stays up through a long agent loop on hardware I already own.
One more practical note: Super’s 1M context is interesting for long agent traces and document-heavy jobs, but memory pressure is still real. Bolt’s 24 GB target is the clearer default for a single workstation GPU. Start there unless you already know you need the longer window.
Update (NVIDIA blog, Oct 15): NVIDIA’s Nemotron 3.5 Super VL post frames Super as an open multimodal / vision-language agent model (document/OCR/chart workflows), with Bolt 74B-A9B for ~24GB local harnesses and a smaller Bolt E4B 4GB edge variant coming Oct 23. Technical write-up: developer blog. Weights via Hugging Face / NIM as described there.
Have a blessed day!