Ollama 0.40.14 focuses less on flashy new checkpoints and more on the behavior developers actually receive when they call a local model through familiar APIs. The official November 3 release notes say its OpenAI-compatible endpoints no longer replace a model's packaged sampling defaults when a client omits temperature or top_p. Previously, the compatibility layer could supply a value of 1.0 instead. That is a small-looking API fix with meaningful implications for reproducibility, because the decoding distribution can change even when the weights, prompt and model name remain identical.
The new behavior is aligned with Ollama's native /api/chat handling: model-provided defaults apply unless the client sends an explicit setting. Applications that have always supplied their own temperature and top_p should not assume their output behavior has changed. But teams using OpenAI SDKs against Ollama may want to rerun reference prompts and inspect configuration logs. Consistency across interfaces can matter more for regression tests and agent workflows than a nominal new feature flag.
The release also fixes decoding of cloud web-search responses on /api/chat, including compressed payloads, and improves how /v1/responses retains earlier readable input content, including previous assistant refusals. Ollama reports faster top-k sampling on MLX for large vocabularies, without a single universal throughput number in the announcement. Consequently, it would be inappropriate to attach a blanket speedup or a new quality ranking to this release.
Other listed changes target developer tools: Claude Code cloud-model launches now honor the relevant output-token limit, the app can retry failed ChatGPT disconnects, and Codex launching checks whether ChatGPT routing is available. Those cloud integrations should not be confused with a new open-weight model or a local-only deployment guarantee. The practical takeaway is to validate model-specific sampler settings, response compatibility and Apple Silicon performance against one's existing workload before treating an Ollama upgrade as a behavior-preserving patch.
What this changes for the evidence layer
Ollama is an open-source runtime, not a new source of model weights. Model-specific licenses and canonical weight revisions remain separate. Official release-notes claims are not independently benchmarked.
Ollama project — official v0.40.14 release notes
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗
