Ollama v0.40.5 adds NVIDIA Nemotron 3.5 Super, a 120B MoE vision-language model with up to 1M-token context

Ollama v0.40.5 adds support for NVIDIA's Nemotron 3.5 Super, a 120B-parameter vision-language model with 12B parameters active per token and context up to 1M tokens. The release also adds a smaller sibling, Nemotron 3.5 Bolt, and ships fixes to the OpenAI-compatible Responses API and cloud error handling.

Nemotron 3.5 Super

According to the Ollama release notes, Nemotron 3.5 Super targets coding, reasoning, tool use, and document understanding. Specifications listed:

• Architecture: hybrid Mamba-2/Transformer mixture-of-experts backbone
• Parameters: 120B total, 12B active per token
• Vision: C-RADIO vision encoder
• Context window: up to 1M tokens
• Features: tool calling and an on/off thinking mode

It runs locally with:

ollama run nemotron-3.5-super

Nemotron 3.5 Bolt

The notes describe Nemotron 3.5 Bolt as a compact sibling that distills Super down to 74B parameters (9B active). Ollama says it is tuned to fit a 24 GB NVIDIA GPU for everyday agentic tasks, and that it includes a drafter model for speculative decoding. The notes do not specify the quantization behind the 24 GB claim, so real-world memory requirements are unconfirmed.

Other changes in v0.40.5

• Cloud decision models: clef:cloud and clef-flash:cloud run decision models through ollama.com cloud without a local pull. The release notes do not say who built these models or give their specifications.
• Responses API fix: Continuing an OpenAI-compatible Responses conversation that includes custom tool calls no longer fails with an error.
• Clearer cloud errors: Messages now explain sign-in problems, payment issues in the desktop app, and upstream access errors from the ChatGPT gateway.
• Dependencies: llama.cpp is updated, adding Nemotron 3.5 support, along with MLX.

The release has one new contributor, @willietran, whose first contribution is PR #18986.

What is not disclosed

The release notes contain no benchmark scores for either Nemotron model, no training cutoff date, and no pricing for the cloud-hosted decision models. They also do not state the licensing terms of the Nemotron 3.5 weights. Those details would need to come from NVIDIA's own documentation or the Ollama library pages.

What this means

This is a runtime and distribution update, not a model launch by itself. Ollama is making NVIDIA's models available to local users, and the llama.cpp update is what enables Nemotron 3.5 support.

The notable part is the combination of a hybrid Mamba-2/Transformer design and 1M-token context. If the architecture performs as described, the Mamba-2 layers should keep long-context memory costs lower than a pure Transformer, though the notes do not quantify this. A 120B MoE with only 12B active parameters still needs enough memory to hold all the weights, so Super is a workstation or multi-GPU model. Bolt is the variant aimed at single-GPU users.

The Responses API fix matters for developers building agents against Ollama's OpenAI-compatible endpoint, where custom tool calls previously broke conversation continuation. Without published benchmarks, the performance claims for both models remain unverified.