builders · review · first seen Nov 3, 10:32 UTC

Ollama fixes cloud search decoding and model defaults

Ollama v0.40.14 shipped fixes for cloud web search decoding, model sampler defaults, Claude Code output limits, and ChatGPT disconnects. The builder decision is whether those fixes touch a workflow you use today.

L0 · primary1 source4 claims checked4 min read

What shipped

The release fixes decoding of cloud web search responses on /api/chat, including compressed responses. This is a response-handling fix, not an announcement of a new search capability. If your application uses cloud web search through /api/chat, this is the change to check first: compare how your client handles a response in the workflow that previously exposed the decoding problem. The release excerpt does not specify which responses were affected or how often the failure occurred. [1]

OpenAI-compatible endpoints now use a model’s packaged sampler defaults when the client omits temperature or top_p. Before this change, the endpoints substituted 1.0 for omitted values; the release says the new behavior matches /api/chat. That makes omission meaningful. A client that leaves either setting out is asking the endpoint to use the model’s own value, rather than the former endpoint default. A client that sends explicit values is outside the behavior described by this fix. [1]

For Claude Code, ollama launch claude now applies the model’s output token limit. The stated reason is to avoid cloud models running into the 128k default cap. I read this as a limit-selection fix for that launch path, not as a claim that every model can produce a 128k-token output. The release gives no model-by-model limits here, so the applicable limit remains something to check in the model you use. [1]

Ollama’s app can now retry ChatGPT disconnect failures instead of leaving the connection in a broken state. The scope matters: the claim is about a retry path in the app after a disconnect failure. It does not say that disconnects have stopped, that retries always succeed, or that a separate client receives the same recovery behavior. [1]

Why it matters if you build

For an /api/chat application that relies on cloud web search, the useful test is the response your own client consumes. Correct decoding, including for compressed responses, addresses a failure at the boundary between the service response and the application’s next step. I would check the resulting content in that workflow before changing downstream handling. The release supports a decoding fix; it does not establish a broader change to search results or their quality. [1]

The sampler change deserves attention even if an OpenAI-compatible client has no setting to update. If that client omits temperature or top_p, its requests now receive the model’s packaged defaults instead of the former 1.0 substitutions. That can change the behavior a builder observes without a change to the client request. Compare omitted settings with explicit settings before treating a difference as a model change. The release provides no measured effect on output quality, latency, or cost. [1]

The two connection fixes call for narrower checks. If you launch Claude Code through ollama launch claude, verify that the model’s output limit is the limit applied in your workflow, rather than assuming the old 128k default cap. If you use Ollama’s app with ChatGPT, exercise recovery after a disconnect failure. Neither claim supplies a success rate, a cost figure, or a guarantee that a retry resolves every failure. [1][2]

How to try it today

1. Use cloud web search through /api/chat and inspect how your client receives the decoded response, including a compressed response if your workflow encounters one.
2. Send an OpenAI-compatible request without temperature or top_p, then compare it with a request that sets those options explicitly.
3. Run ollama launch claude with the model you use in Claude Code and check the output limit applied to that workflow.
4. In Ollama’s app, check whether a ChatGPT disconnect failure offers a retry and whether your connection recovers.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   Ollama v0.40.14 decodes compressed cloud web search responses on /api/chat correctly.

   per GitHub releases · ollama/ollama (L0)
2. shipped L0 · primary

   OpenAI-compatible endpoints use a model’s packaged sampler defaults when clients send no temperature or top_p.

   per GitHub releases · ollama/ollama (L0)
3. shipped L0 · primary

   ollama launch claude applies the model’s output token limit for Claude Code instead of the 128k default cap.

   per GitHub releases · ollama/ollama (L0)
4. shipped L0 · primary

   Ollama’s app can retry ChatGPT disconnect failures.

   per GitHub releases · ollama/ollama (L0)

What is not confirmed

The release does not quantify how often compressed cloud search responses failed, which clients encountered them, or whether the fix changes anything beyond correct decoding on /api/chat. It also does not measure the output difference produced by using packaged sampler defaults, or provide latency and cost comparisons. Those are questions for a check against your own requests, not conclusions from these notes. [1][2]

The notes do not list the output token limit of each model used with Claude Code. They also do not say that every ChatGPT disconnect retry succeeds. Do not turn either fix into a blanket reliability or capacity claim. [1][2]

Who should act, who should wait

Act now if your application uses cloud web search on /api/chat, if an OpenAI-compatible client omits temperature or top_p, or if your Claude Code workflow uses ollama launch claude. Each case has a specific behavior you can check against your existing request or model. Keep the test narrow: confirm the response, the selected sampler values, or the applied output limit before changing application logic. [1][2][3]

If you depend on Ollama’s app connection to ChatGPT, test its retry path after a disconnect failure. Wait before making a broader rollout decision if you need measured reliability, cost, or output comparisons: none are established by these claims. If none of the named endpoints, launch paths, or app connection applies to you, these notes alone do not create an immediate action. [1][2][3][4]

Sources

• v0.40.14 — GitHub releases · ollama/ollama L0 · primary