Ollama v0.31.2: Agent Core Framework & GPU Flash Attention
Decision Brief
What changedOllama v0.31.2 introduces an Agent core framework and enables Flash Attention for CUDA CC 6.x GPUs, along with various fixes.
Why it mattersAgent framework enables local multi-step Agent workflows, while Flash Attention boosts inference efficiency for older NVIDIA GPUs, benefiting teams deploying locally with legacy hardware.
Who should careOpen-source model users, Inference / infra teams
Affected stackOllamaNVIDIA
Builder actionUpgrade Ollama to v0.31.2
Source confidenceHigh · Official release / blog / repo
Ollama v0.31.2 adds a new Agent core framework, laying the foundation for future Agent capabilities. It also enables Flash Attention on CUDA CC 6.x GPUs (e.g., Pascal architecture), significantly accelerating attention computation. The release fixes CUDA toolkit detection, falls back to standard CUDA when JetPack is missing, and removes unsupported devices for ROCm. The MLX backend has been rewritten and updated. For developers running open-source models locally, Flash Attention on older GPUs notably speeds up inference, while the Agent framework prepares for building automated workflows.
Summary basis: official / RSS sourceCompiled from the source scope noted above; the original remains authoritative.
Sources
- Ollama（GitHub Releases）
Local-model runtime releases: new supported models and serving features.
- Ollama（GitHub Releases）