Ollama: Agent Reliability Fixes and Offline Model Transfer
Today's activity centers on stopping silent failures in agent workflows and easing model portability, with fixes for lost reasoning and broken tool calls alongside new export-import commands and caller-requested log probabilities.
Duration: PT2M8S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-22T13:01:49Z
- Audio duration: PT2M8S
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 Tuesday, September 22nd, 2026.
The lead signal today is reliability for real-world agent workloads — changes that keep reasoning, tool calls, and structured output from failing quietly.
First, compatibility. In P R 18570, the Open A I compatible endpoint now accepts reasoning content as an alias for reasoning. Previously, messages written to DeepSeek's documented contract lost all replayed reasoning with no error, which breaks follow-up requests when tools are present. Separately, in P R 18571, the…
Second, new capability for scoring and moving models. P R 18580 proposes log probabilities for caller-named tokens, so you can score a fixed candidate set in one forward pass even when those tokens aren't in the top guesses. And P R 18578 adds export and import commands plus server endpoints, addressing painful…
Third, engine maintenance and platform polish. There's a Llama C P P version update in P R 18577 and an M L X version bump in P R 18576. On Windows, P R 18572 fixes unmatched wildcards in from lines, and P R 18582 suppresses pop-up consoles during integration discovery. Related fixes improve error messages for…
What's next: watch the discussion on the new log…
Nearby episodes from Ollama
- Thinking Separation Fix and GPU Visibility
- Weekly Recap - MLX Graduation, Reasoning Fixes, and Onboarding Polish
- Compatibility Fixes and Mac Efficiency
- Closing the Gaps Between Model Metadata and Client Compatibility
- Hardening Pass Across Network, Memory, and Release Paths
- The MLX Engine Comes Home
- Streaming Reliability and Parser Cleanup
- MLX Comes of Age, Cancellation Gets Fixed