Ollama: Decision Models Go Usable, MLX Speeds Up on CUDA
Decision models became runnable from the command line with corrected scoring and new integration tests, while the MLX backend gained CUDA kernel ports and faster Gemma 4 prefill. Cloud billing saw stopgap fixes for usage credits and payment errors.
Duration: PT2M28S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-10-14T13:01:01Z
- Audio duration: PT2M28S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning, it's October 14th, 2026. You're listening to Ollama.
Our lead story: decision models moved from experimental to practical, with command-line support, a scoring correctness fix, and heavier testing landing together.
First, usability. In PR 18962, `ollama run` now works with decision models like Nimble and Clef, which previously failed saying the model does not support generate. You can ask a yes-or-no question with the ask option, or pick from labels with the choice option, or score on levels with the level option. State comes…
That matters only if scoring is right. PR 18974 fixes MLX decision models scoring JSON states differently than the same models on GGUF, by up to point zero eight five. Ollama had compacted JSON before building prompts, while publishers and llama dot cpp use Python-style spacing. It also fixes Tev One, which was…
Second, the MLX backend on CUDA. PR 18657 ports Metal-only custom kernels for Mamba-style scan, convolution, and gated delta operations to CUDA, avoiding slow graph fallbacks. PR 18809 switches Gemma 4 prefill with wide head dimensions to MLX's S D P A, reporting about twelve times faster prompt processing on the…
Finally, cloud use. PR…
Nearby episodes from Ollama
- Proxy Transfers and API Resilience
- Model Compatibility Catch-Up and Windows Reliability
- Responses Compatibility and Clearer Errors
- Decision Scoring Matures as Inference Hardens
- Native Decision Models and Precision Fixes
- Weekly Recap - Decision Models Go Native, Embeddings Go Multimodal
- GPU Discovery Fallback and Docs Cleanup
- OpenAI Parity and Reasoning Fixes