disable reasoning per request llama.cpp
The story begins on April 15th, when a user on LocalLLaMA asked whether llama.cpp's server could disable reasoning for specific requests while keeping it enabled by default. A reasonable question — if you're running a general-purpose inference server, not every query needs a model to think out loud before answering, and the overhead of extended reasoning chains adds latency and tokens you sometimes simply don't want. The post scored modestly and drew some attention, but no definitive answer apparently landed. Then on May 19th, over a month later, someone else arrived at the same wall and asked nearly the identical question. That's the tell. When the same gap surfaces twice in five weeks from different users, it points to a genuine missing capability rather than a one-off edge case.
If confirmed, here is what this means. Per-request reasoning control would be a quiet but genuinely useful feature for anyone deploying reasoning-capable models in production through llama.cpp's server. Right now, if reasoning is on by default, every request — whether it's a simple lookup or a complex multi-step problem — potentially goes through the same extended thinking pipeline. Giving developers a request-level toggle means they can route cheap queries cheaply and expensive queries appropriately, without spinning up separate server instances with different configurations. For latency-sensitive applications, that matters. The second-order effect is that it normalises a tiered approach to inference — the idea that a single model can operate in different cognitive modes depending on context, which is going to become increasingly standard as reasoning models proliferate.
Watch for a pull request or GitHub issue in the llama.cpp repository addressing per-request reasoning toggles — that would confirm this is moving from community request to active development.
<!-- SC_OFF --><div class="md"><p>I run the same model: `google_gemma-4-E2B-it-IQ3_M.gguf` with lmstudio or llama-server</p> <p>and I connect thru `/v1/chat/completions` EP.</p>
<!-- SC_OFF --><div class="md"><p>Even the trick of editing the reply with the tag <think> or <|think|> doesn't do anything for me. On some models I used to be able to
<!-- SC_OFF --><div class="md"><p>Is there an easy way to know if a model is using CPU/RAM (and not only GPU/VRAM)?<br /> (I think standard verbose output, which got shorter, says
<!-- SC_OFF --><div class="md"><p>currently in order to use MTP one needs to enable it in the starting argument of llama server.</p> <p>--spec-type draft-mtp<br />
<!-- SC_OFF --><div class="md"><p>I run ollama in combination with LibreChat using docker compose. Have been using gemma3 for quite some time. Now switched to gemma4 only to
<!-- SC_OFF --><div class="md"><p>I have been using llama.cpp to run some models recently. For example, I've been running GLM-4.7-Flash with this command <code>.\llama-server.exe
<!-- SC_OFF --><div class="md"><blockquote> <p>Wait, I need to be careful about the "no_think" tag in the system prompt. The system prompt says <code>/no_think</code>.
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →