Modelpublic

Piecrust/Spike-4B-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes85downloads

Model Card

Spike-4B · GGUF

Spike is the on-device assistant in the Spike AI iOS app; this GGUF build is for CPU/GPU serving via llama.cpp (a current build with qwen35 support). 📱 Get it on the App Store: https://apps.apple.com/app/spike-ai/id6749781844

A LoRA fine-tune of Qwen/Qwen3.5-4B (a vision-language model), specialized for Spike's tool-calling — reminders, calendar, Apple Home, maps, web, files, code, and the SSH/agent toolset, while staying a natural conversationalist. English + German. Tool grammar: tool:<name> {json}.

Files

Spike-4B-Q4_K_M.gguf (≈2.7 GB) + Spike-4B-F16.gguf (≈8.4 GB source).

Qwen3.5 is a new hybrid (linear-attention + full-attention) architecture; for on-device iOS the app ships the MLX build.

Eval — Spike harness (base Qwen3.5-4B → Spike-4B)

Metric	Base	Spike-4B
Tool calls · thinking-off	42.4%	99.8%
Tool calls · thinking-on	—	99.8%
Vision (image → tool / answer)	68.1%	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.

The Vision row is the fine-tune's image → tool / answer ability, exposed by the MLX build; this GGUF is text-only (tools + thinking).

Usage

—Trained on Spike's compact system prompt; use that exact prompt.
—Optional reasoning via the enable_thinking chat-template kwarg.
—One text tool call per turn: tool:<name> {json}.

License

Derivative of Qwen3.5-4B under the Apache 2.0 License.