---
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
library_name: mlx
pipeline_tag: text-generation
tags:
  - mlx
  - vision-language
  - tool-calling
  - function-calling
  - spike
language:
  - en
  - de
thumbnail: https://huggingface.co/Piecrust/Spike-9B-MLX/resolve/main/banner.png
---

<p align="center">
  <img src="https://huggingface.co/Piecrust/Spike-9B-MLX/resolve/main/banner.png" alt="Spike-9B-MLX" width="100%">
</p>

# Spike-9B · MLX (4-bit)

**Spike** is the assistant in the **Spike AI** iOS app. This is the **largest Spike tool model** — 4-bit MLX for Apple silicon, served via `mlx-swift` / `mlx-vlm`.
📱 **Get it on the App Store:** https://apps.apple.com/app/spike-ai/id6749781844

A LoRA fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (a vision-language
model), specialized for Spike's tool-calling — reminders, calendar, Apple Home, maps, web, files,
code, and the SSH/agent toolset — **plus vision** (flyer → calendar, note → reminder, receipt → answer),
while staying a natural conversationalist. English + German. Tool grammar: `tool:<name> {json}`.

## Files
4-bit MLX weights (`model.safetensors`, ≈5.8 GB) + tokenizer, processor, chat template. Load with `mlx-swift` / `mlx-vlm` on Apple silicon.

> Qwen3.5 is a new *hybrid* (linear-attention + full-attention) architecture; a text-only [GGUF build](https://huggingface.co/Piecrust/Spike-9B-GGUF) is also available for llama.cpp servers.
> This is a **server-class 9B** — it targets Macs / workstations, not phones (the on-device app ships the 2B / 4B builds).

## Eval — Spike harness (base Qwen3.5-9B → Spike-9B)

| Metric | Base | **Spike-9B** |
|---|---:|---:|
| Tool calls · thinking-off | 52.0% | **99.8%** |
| Tool calls · thinking-on | — | **99.8%** |
| Vision (image → tool / answer) | 72.5% | **100%** |
| Normal-chat tool-leak (lower=better) | 1.6% | **0%** |

Trained text+thinking+German, then a vision-replay stage, then a **conversation-repair** stage
(distilled base-model chat + contrastive tool/vision replay) so it keeps `enable_thinking`
reasoning and vision, tool-calls at 99.8%, **and does not hijack casual chat into tool calls**.

## Usage notes
- Trained on Spike's **compact system prompt**; use that exact prompt for best results.
- Supports optional reasoning via the `enable_thinking` chat-template kwarg.
- Vision: pass an image with the user turn; the model reads it and answers or calls a tool.
- Tool calls are plain text `tool:<name> {json}` — one per turn.

## License
Derivative of Qwen3.5-9B under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).