Ollama: Weekly Recap - Decision Models Go Native, Embeddings Go Multimodal
Ollama shifted decision-model scoring into upstream llama.cpp and added multimodal embeddings, alongside fixes for OpenAI compatibility and Windows model storage. The pattern this week is retiring custom patches in favor of native runner support.
Duration: PT3M45S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-10-12T09:08:49Z
- Audio duration: PT3M45S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Hello, this is your Ollama briefing for the week of October 5th through October 12th.
This week: 50 pull request activity items, 28 additional commits this week.
The lead insight: Ollama is handing off specialized work to its runners. Decision-model scoring moves into llama dot cpp, multimodal embeddings arrive on both major back ends, and a wave of compatibility fixes cleans up what is left behind.
First, decision models go native. In PR 18917, Ollama updates llama dot cpp to build 11521, which adds built-in support for the System One endpoint and EmbeddingGemma 2. That lets GGUF decision models send scoring requests straight to llama dot cpp, removing Ollama's own scoring server and the Clef patch. Supporting…
Second, embeddings become multimodal. PR 18820 implements the EmbeddingGemma 2 architecture on the M-L-X runner, with a bidirectional text encoder and shared vision and audio towers, and allows the embed API to accept per-item media. A companion commit adds multimodal embedding support through the llama server path.…
Third, compatibility hardening across APIs, storage, and launch. On the OpenAI side, PR 18722 keeps tool message parts in one message to preserve call…
Nearby episodes from Ollama
- Native Decision Models and Precision Fixes
- GPU Discovery Fallback and Docs Cleanup
- OpenAI Parity and Reasoning Fixes
- Legacy Model Cleanup and Agent Launcher Polish
- Memory Accuracy, Tool Fixes, and Simpler Onboarding
- Multimodal Embeddings and Cloud Convergence
- MLX Performance Push and Tool-Call Reliability
- Tokenizer Parity and RC-Aware Updates