---
title: Salience Chat
emoji: 🌿
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 5.38.0
app_file: app.py
pinned: false
license: apache-2.0
models:
  - vectionlabs/Salience-27B-R5
short_description: Salience 27B R5 — reasoning that stops when done
---

# Salience Chat

A minimal chat interface for [vectionlabs/Salience-27B-R5](https://huggingface.co/vectionlabs/Salience-27B-R5),
loaded in **NF4** on **ZeroGPU**.

- **Model**: Salience 27B R5 — 27.8B dense, every parameter active on every
  token, 1,048,576-token context, Apache-2.0
- **Reasoning**: **native**. The effort dial sends the model's own
  `reasoning_effort` template parameter, not prompt text
- **Stack**: React (Vite) frontend served as static files; a Gradio app mounted
  at `/gradio` provides the streaming API consumed via `@gradio/client`

## What changed moving from 9B to 27B

**NF4, not bfloat16.** At 27.8B, bf16 would be 51.7 GiB and no ZeroGPU tier
holds it. NF4 quantises the body to ~11.8 GiB and leaves the embedding and
output head in bf16 — bitsandbytes skips `nn.Embedding` and `lm_head` by
design — landing around 16.5 GiB.

**The prompt scaffolding is gone.** The 9B build forced reasoning by pasting
instructions into a system prompt (*"reason step by step"*, *"write a rough
draft, check it"*) and prefilling an opening `<think>` tag, because that
model's template had no thinking support.

R5's does. It takes a real `reasoning_effort` parameter and emits the opening
tag itself, so all of that text was deleted rather than adapted. Keeping it
would have been actively harmful: instructing a natively-reasoning model to
reason makes it *perform* reasoning for the reader instead of doing it, and
the two compound into narrated pseudo-thought.

## The effort dial

R5's template accepts exactly three values. The UI ships four pills, so High
and xHigh both send `xhigh` and differ only in how long they may run —
inventing a fourth setting that does nothing would be theatre.

| pill | sends | token budget |
|---|---|---|
| Low | `reasoning_effort: low` | 2,048 |
| Medium | `reasoning_effort: medium` — the model's own default | 4,096 |
| High | `reasoning_effort: xhigh` | 8,192 |
| xHigh | `reasoning_effort: xhigh` | 16,384 |

`medium` injects no deliberation instruction at all: the model decides. That is
R5's headline behaviour and it is what this Space is on by default.

## The `<think>` trap, handled server-side

The chat template emits the **opening** `<think>` as part of the generation
prompt. With `skip_prompt=True` the stream therefore starts *inside* the block
and the completion carries only the closing tag. Any client looking for a
matching pair finds none, reports zero thinking, and renders the whole chain of
thought as the answer. `app.py` re-adds the opening tag to the stream so the
frontend's parser sees a well-formed pair.

## Why Gradio SDK and not Static

ZeroGPU is only available on Gradio-SDK Spaces. A Static (React template) Space
cannot run a local model. This repo keeps the React frontend but ships it as a
pre-built bundle (`static/`) served by the same process that hosts the
ZeroGPU-backed Gradio API.

## Local frontend development

```bash
cd frontend
npm install
npm run dev     # dev server, proxies /gradio to a running app.py
npm run build   # emits ../static — commit this
```