Ollama: Single-Pass Structured Outputs and MLX Speedups

Thinking models with structured outputs move from two generations to a single pass, fixing dropped content, metric stitching, and stray tokens. MLX prompt processing gets 14 to 19 percent faster and Gemma 4 vision uses dynamic resolution, alongside fixes for model lookup and container GPU support.

Duration: PT2M26S

Episode overview

This episode is a short developer briefing from Ollama.

It explains recent repository work in plain language.

  • Show: Ollama
  • Published: 2026-09-23T13:03:40Z
  • Audio duration: PT2M26S

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 September 23rd, 2026, and this is your Ollama developer briefing.

The headline is a fix for structured outputs on thinking models, moving from two generations to a single pass, alongside a focused push to speed up the M-L-X stack.

First, thinking models with a format. Until now, the server ran an unconstrained generation, canceled it once content appeared, then re-ran the prompt under a grammar. As described in P-R 18479, that cost a second prefill, dropped the chunk crossing the boundary, required a Harmony prompt workaround, stitched…

Second, M-L-X performance and vision. P-R 18550 uses the gated delta kernel and folds scaling into the activation to lift Qwen prompt processing roughly 14 to 19 percent on M-5 Max, enabled by the version bump in P-R 18576. Separately, P-R 18603 replaces the fixed image budget for Gemma 4 with per-image selection…

Finally, reliability and compatibility. Merged P-R 18438 fixes intermittent model-not-found errors where case-insensitive name matching across thousands of manifests plus randomized map ordering could overwrite tag casing. Merged fixes also stop the desktop app from hanging when checking for Chat-G-P-T or Codex, and…

Open…

Nearby episodes from Ollama

  1. Agent Reliability Fixes and Offline Model Transfer
  2. Thinking Separation Fix and GPU Visibility
  3. Weekly Recap - MLX Graduation, Reasoning Fixes, and Onboarding Polish
  4. Compatibility Fixes and Mac Efficiency
  5. Closing the Gaps Between Model Metadata and Client Compatibility
  6. Hardening Pass Across Network, Memory, and Release Paths
  7. The MLX Engine Comes Home
  8. Streaming Reliability and Parser Cleanup