Ollama

Recently shipped NVIDIA Nemotron 3.5 Super and Bolt to the model library, plus cloud-hosted decision models.

Nemotron 3.5 lands in two sizes.

• Super pairs a 120B hybrid Mamba-2/Transformer mixture-of-experts backbone (12B active per token) with a C-RADIO vision encoder, plus tool calling, an on/off thinking mode, and context up to 1M tokens.
• Bolt distills it to 74B (9B active) tuned for a 24 GB NVIDIA GPU, with speculative decoding through an included drafter.

Both cover vision-language work: coding, reasoning, document understanding.

Decision models now run from the cloud. clef:cloud and clef-flash:cloud serve structured decisions through ollama.com without pulling weights locally. Typed questions still return choices, probabilities, and scores for triage, routing, and classification, and Clef's multimodal scoring accepts images alongside the text state.

The Apple Silicon MLX path widened. Supported architectures keep running on MLX by default, the Metal custom kernels gained a CUDA port, and PPLX Decider v1.2 is supported. GGUF also picked up the q2_0 tensor and file type.

Launch and interop cleanups. ollama launch claude and ollama launch codex no longer warn about Claude connectors or service tiers, and installed models appear in the Codex /model picker. Continuing an OpenAI-compatible Responses conversation with custom tool calls no longer fails. Background model upgrades are paused, cutting overhead on every request in high-volume use such as embeddings.