Ollama released version v0.40.0-rc6, which includes multimodal embeddings. This update adds support for per-item media via input dictionaries. The release is aimed at developers and users interested in advanced AI features.
Key facts
| Fact | Detail | The source says |
|---|---|---|
| Architecture | EmbeddingGemma2Model | “Implements the EmbeddingGemma2Model architecture on the MLX runner” |
| Text Encoder | 24-layer bidirectional text encoder with PLE | “24-layer bidirectional text encoder with PLE” |
| Vision/Audio Towers | shared gemma4 vision/audio towers | “shared gemma4 vision/audio towers” |
| Output | mean-pool + L2 output | “mean-pool + L2 output” |
What happened
Ollama has released version v0.40.0-rc6, which includes multimodal embeddings. This update implements the EmbeddingGemma2Model architecture on the MLX runner, featuring a 24-layer bidirectional text encoder with PLE, shared gemma4 vision/audio towers, and mean-pool + L2 output. The /api/embed endpoint now accepts per-item media via input dictionaries. This release is aimed at developers and users interested in advanced AI features.
In the Fathom Layer index
- Ollama — Developer Tools
Source: Ollama
