Ollama + `qwen3:4b`: `think: false` and `/no_think` don't turn thinking off; answers come back empty (`done_reason: length`) or with the reasoning mixed into `response`

Context

Source(s): https://juejin.cn/post/7697205221650972687 (the author's own measured write-up with raw JSON responses); related card: fix/13 (empty content when max_tokens is too small for a qwen reasoning model) Status: reported working by source, not reproduced by us.

Problem

In a local RAG app, the last step asks qwen3:4b on Ollama for a short answer. Following the docs, the author set think: false, and the answer came back as an empty string with done_reason: length. With a bigger budget and think: false, the thinking field became empty and the full reasoning text (about 2,100 characters, ending in a bare </think> tag) was put in response ahead of the answer. Putting /no_think at the start of the prompt did not help either: the model treated it as part of the question and kept reasoning until the budget ran out. Through the OpenAI-compatible endpoint (the author warns that Spring AI clients hit this), the app got HTTP 200 with an empty content, because the reasoning sits in a non-standard reasoning field.

Error (redacted / normalized)

{ "model": "qwen3:4b", "response": "", "done_reason": "length", "eval_count": 200 }

OpenAI-compatible endpoint: content = "" (0 chars), reasoning = <full reasoning>, finish_reason = "length", HTTP 200

"thinking": "", "response": "嗯，用户让我用一句话说明……</think> <the actual answer>"

Applies when

Windows 11, RTX 4060 Ti 8GB, Ollama 0.40.3, qwen3:4b (Q4_K_M), called on /api/generate, /api/chat and /v1/chat/completions. Oct 2026.

Confirm it is this problem

1. A response has done_reason (or finish_reason) length and an empty response / content → expect: the output budget was used up by reasoning (eval_count or completion_tokens equals your budget).
2. The answer text starts with the model talking to itself ("嗯，用户让我……", "I need to confirm…") or contains </think> → expect: you have think: false set and the reasoning has moved into the answer field.
3. Output token count is far larger than the answer: the author measured a 30–34 character answer costing about 1,150–1,580 tokens and 13.8–19 seconds, while plain generation ran at 88 tokens/s.

Not this when / boundary

If the answer is empty but the reasoning field is also empty and done_reason is stop, this card doesn't apply. If you're not using Ollama (for example vLLM or a hosted qwen API) and only see empty content with a small max_tokens, see fix/13 first. Newer Ollama versions may honor think: false for this model; this may change between versions (our inference).

Root cause

On Ollama 0.40.3, qwen3:4b is a hybrid reasoning model, and in the author's tests the think switch did not stop it from reasoning: think: false only stops Ollama from separating the reasoning into the thinking field, so the reasoning lands in response instead. /no_think in the prompt was not honored on this path. Each call generated roughly 1,150–1,580 tokens even for one-sentence questions, almost all of it reasoning, and the author found reasoning length had little to do with question difficulty, so small budgets (200, and 700 in the author's test) end before any answer is written.

Variables

<budget>: options.num_predict on the native endpoints, or max_tokens on /v1/chat/completions.

Fix

1. Remove think: false from your requests. → expect: the reasoning goes back to thinking (native endpoints) or reasoning (OpenAI-compatible endpoint), and the answer field holds only the answer.
2. Raise <budget> to at least 2048 and read only response (native) or message.content (OpenAI-compatible), ignoring the reasoning field. → expect: done_reason: stop and a non-empty answer. The author measured this on /api/generate (30-character answer, 13.8 s) and /v1/chat/completions (34-character answer, 18.1 s) and on a real RAG prompt (correct answer, 19 s).
3. If you don't need reasoning (for example, short answers grounded in retrieved text), switch to the non-thinking variant of the same size: ollama pull qwen3:4b-instruct, then use qwen3:4b-instruct as the model name. → expect: answers without a reasoning phase. This is the author's recommended fix, but the saved article text shows no measured output for it.
4. Always pass num_ctx explicitly (the author used num_ctx: 4096); the model reports a 262,144-token maximum, and the author warns that opening a KV cache that large overflows an 8GB card.

How to verify

Send the same question with num_predict 2048 (or max_tokens 2048) → expect: done_reason/finish_reason stop and a short answer in response/content with no reasoning text in it.

What did not work

think: false on /api/generate and /api/chat, /no_think at the start of the prompt, switching between the native and OpenAI-compatible endpoints, and raising the budget to 700 all still gave empty answers or reasoning mixed into the answer.

Rollback

Switch the model name back to qwen3:4b and restore your previous budget.

Risks

A 2048-token budget on the thinking model means 10–19 seconds per answer on the author's hardware, so tell users to expect that delay. Clients that only read content (Spring AI per the author; likely other OpenAI-compatible SDKs too, our inference) get a successful-looking empty string when the budget runs out, so check for an empty answer before showing it. Don't try to strip the reasoning by cutting the text at </think> (our inference: the author shows the reasoning and the answer mixed in one field with only a bare model-generated tag between them, so there is nothing reliable to filter on).

Verified scope

Ollama 0.40.3 with `qwen3:4b` (Q4_K_M), Oct 2026.

Becomes obsolete when

An Ollama release makes think: false actually suppress reasoning for qwen3 models.