Ollama: MLX Tokenizer Fixes and Thinking Model Reliability
MLX dominates the day with tokenizer correctness fixes, new SystemOne and Kolibri 1 support, and decision-model tuning. Parallel fixes tighten structured output for thinking models and harden Windows, GPU discovery, and API handling.
Duration: PT2M41S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-10-04T13:00:52Z
- Audio duration: PT2M41S
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 October 4th, 2026, and this is your Ollama developer briefing.
The lead signal today is correctness for local models — matching publisher behavior for tokenizers and for thinking models, so the same prompt produces the same IDs and the same structured output across backends.
First, the MLX stack. Pull request 18779 fixes token ID mismatches by restoring pretokenizer stages, Unicode and whitespace boundaries, added-token handling around normalization, and ranked BPE merges, with consistent handling for empty tokens. It's backed by a regression table across eight published MLX models,…
Second, thinking models and structured output. With thinking enabled, Gemma 4 could answer directly with a scalar like 391 despite an object schema. Pull request 18783 carries parser boundaries through both API routes and requires a thinking close before the final format, keeping direct answers constrained in both…
Third, reliability hardening. On Windows, System One requests to Clef Flash failed with non-finite logits because head-weight reads past 2 gigabytes truncated the seek offset to 32 bits — pull request 18777 fixes that. Pull request 18773 stops zero-memory pseudo-devices…
What'…
Nearby episodes from Ollama
- Tool Call Ordering Fixes and Decision Models
- Decision Models Take Shape
- Pointer-Head Scoring and Reliability Fixes
- System One Moves to Explicit Capabilities
- System One Decisions Launch
- Silent Sampling Overrides Fixed
- Weekly Recap - Single-Pass Structured Outputs and MLX Speedups
- Tool-Call Reliability Fixes