Ollama v0.40.14 Aligns Sampler Defaults and Enforces Output Limits for Claude Code

What changed#

Ollama v0.40.14 delivers several targeted fixes across API handling, command-line tooling, and backend execution:

• Output token caps for Claude Code: The command ollama launch claude now uses the model’s explicit output token limit, ensuring cloud models are no longer constrained by the previous 128k default cap on outputs.
• OpenAI endpoint sampler parity: OpenAI-compatible API routes no longer fall back to 1.0 for missing temperature or top_p values. Requests omitting these parameters now inherit the model’s packaged default samplers, matching the native /api/chat behavior.
• Payload decoding and context preservation: Cloud web search responses returned via /api/chat are now correctly decoded, including compressed payloads. The /v1/responses route now preserves readable input structure and past assistant refusals instead of flattening the prompt history.
• Routing and stability: The Ollama desktop app automatically retries disconnected ChatGPT sessions. Additionally, ollama launch codex checks ChatGPT routing prior to setting up configuration and returns explicit error messaging if the route is offline.
• MLX sampling speedups: Top-k sampling performance on MLX backends has been optimized for models with large vocabulary sizes.

Why it matters#

Implicit parameter overrides can lead to unexpected output variance. Previously, clients sending requests without explicit sampler arguments to OpenAI-compatible endpoints received forced 1.0 values, overriding model-specific tuning. Standardizing parameter resolution across both native and OpenAI endpoints provides predictable generation outputs.

Respecting model output limits in ollama launch claude prevents long generation tasks from terminating prematurely against an artificial 128k default cap. This fix is particularly relevant for agentic developer workflows running on Claude Code. Meanwhile, maintaining unflattened context and explicit refusal history in /v1/responses prevents conversation context degradation across complex multi-turn API calls.

What to do next#

Update your local deployment to Ollama v0.40.14. If your integration relies on OpenAI-compatible endpoints and expects an implicit 1.0 default for temperature or top_p, update your client request bodies to send those values explicitly. Otherwise, your requests will now default to the underlying model’s embedded parameters.

Sources#

• Ollama v0.40.14 Release Notes