This model is a LoRA fine-tune of Qwen/Qwen3.5-2B, specialized for Spike's tool-calling workload: reminders, calendar, Apple Home, maps, web search/browse, files, code, and the full SSH/agent toolset — in English and German. It emits Spike's text tool grammar:

tool:<name> {"key":"value"}

Note on the GGUF: Qwen3.5 is a new hybrid (linear-attention + full-attention) architecture. Loading requires a current llama.cpp build with qwen35 support. For on-device iOS the app ships the MLX build instead.

Files

File	Use
`Spike-2B-Q4_K_M.gguf` (1.2 GB)	Recommended quant for CPU/GPU serving
`Spike-2B-F16.gguf` (3.8 GB)	Full-precision source for requantizing

Eval — Spike harness (before → after)

Metric	Base Qwen3.5-2B	Spike-2B
Tool calls (thinking off)	39.8%	99.6%
Tool calls (thinking on)	—	96.8%
Vision (image → tool / answer)	67.5%	100%
Valid JSON on tool calls	≈64%	100%
Normal-chat tool-leak (lower=better)	—	0%

Trained in three stages — text + thinking + general + German (≈22.5k samples), an 800-image vision-replay stage (flyer → calendar, note → reminder, receipt/scene → answer), then a conversation-repair stage — so the model keeps its enable_thinking reasoning and vision, speaks Spike's tool grammar, and does not hijack casual chat into tool calls (normal-chat tool-leak 0%). (Vision is available in the MLX build; the GGUF text model carries tools + thinking.)

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.
Tool calls are plain text tool:<name> {json} — one per turn.

License

Derivative of Qwen3.5-2B under the Apache 2.0 License.

Piecrust/Spike-2B-GGUF

Spike-2B · GGUF

Files

Eval — Spike harness (before → after)

Usage notes

License

Piecrust/
Spike-2B-GGUF