Ollama: Model Behavior Bugs and Blob Security Hardening
Today's activity centers on fixing model-family misdetection that broke thinking and format controls, alongside a new security patch closing off unsafe blob redirects during downloads. Together they show ongoing cleanup of edge cases in how Ollama identifies and handles specific model architectures.
Duration: PT2M42S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-08T13:01:09Z
- Audio duration: PT2M42S
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 July 8th, and here's what moved in Ollama's codebase.
The clearest pattern today is model identification bugs quietly breaking user-facing behavior. PR 15901 generalizes a fix that previously only applied to Gemma 4: when a user sets think equals false, Ollama now correctly applies format constraints across all thinking-capable models, including Qwen 3.5 and Qwen 3…
Second theme: memory and hardware fitting got more precise. PR 16996 lets integrated GPUs — specifically GB10 and Strix Halo systems — offload the multimodal projector instead of forcing it onto CPU. Ollama now pads its internal fit target to account for the projector's memory needs, fixing unnecessary CPU fallback…
Third, security and safety hardening. PR 17082 rejects blob redirects that could route registry downloads to local or private network targets, closing a DNS-rebinding style risk. PR 17084 caps static template ranges so a malicious or malformed model template can't force excessive CPU-bound rendering.
Smaller housekeeping: PR 17071 removes dead internal code, and 17079 and 17076 clean up install and Docker documentation for Jetson and Intel GPU users.
What to remember: if you're testing…
Nearby episodes from Ollama
- Launch Gets Smarter, Model Support Gets Wider
- Cleaning Up the Scheduler's Edge Cases
- Weekly Recap - Rebuilding the Model Pipeline and Tightening the Guardrails
- Truth in Reporting
- MLX Create Pipeline Rewrite Lands
- Agent Harness Lands, Hardware Support Gets a Cleanup
- Gemma 4 Support and Platform Improvements
- Weekly Recap - MLX Performance & Path Handling