README.md

---
title: Does this model need a reasoning parser?
emoji: 🧩
colorFrom: indigo
colorTo: gray
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: Decide --reasoning-parser from a model's own chat template
---

# Reasoning-parser advisor

A model that thinks needs vLLM's `--reasoning-parser` to split that
thinking out of the answer, and the two ways of getting it wrong are not
symmetric.

A **missing** parser is visible: the user sees the model's reasoning in
the reply. A **wrong** one is not — it claims the entire output, the
request finishes with `stop`, and `content` comes back **empty**.
Nothing errors, and nothing in the logs says so.

The family name cannot decide it. `Qwen3.6-35B-A3B` emits `<think>` and
needs the `qwen3` parser; `Qwen3-VL-30B` emits nothing of the sort and
is broken by it. A quantized pack inherits its source's template, so the
same split runs through every repack.

Everything runs **in the visitor's browser**. There is no backend and
nothing is uploaded: the page fetches the repo's chat template — a few
kilobytes — and asks whether the assistant's turn carries a marker some
parser closes on.

## How the rules are kept honest

Markers are read out of vLLM's own `vllm/reasoning/` modules rather than
from documentation, because a parser matches what its code matches. Only
markers that *discriminate a reasoning section* are listed: an earlier
version of the table included `<|end|>`, gpt-oss's turn terminator, and
recommended a parser for a model that does no thinking at all.

The rules exist twice — once in Python for development, once in
JavaScript for this page — and a parity gate runs the same fixtures
through both and refuses to deploy on any disagreement.

Built from the tooling behind the [aleada](https://huggingface.co/aleada)
W4A16 packs. See also the
[pack integrity checker](https://huggingface.co/spaces/aleada/pack-integrity-check).