Ollama v0.40.4 RC adds decision models and EmbeddingGemma 2, shows local models in Codex model picker

Verified (fact)

Ollama released v0.40.4 release candidates (rc1 and rc2), which update llama.cpp to b11521 to support decision models and EmbeddingGemma 2. On the MLX side, it adds Kev decision-model scoring and recurrent state compaction that is restored after speculative decoding rollback. Installed local models now appear in Codex's /model picker, and the inference path reuses verified manifests and models that are already loaded.

Why it matters (interpretation)

Decision models and multimodal embeddings become available right away in local environments, and the Codex integration is stronger.

Sources

• Official v0.40.4 Ollama
• Official v0.40.4-rc1: Add Kev decision scoring to MLX (#18976) Ollama

More about Ollama

• 2026-10-07 P2 Open source Generally available Ollama
  Ollama v0.40.0 makes the MLX runtime the default on Apple Silicon

  Starting with Ollama v0.40.0, model architectures supported by the MLX runtime run on MLX automatically on Apple Silicon devices....