Qwen3.5-2B — GGUF (iPhone-optimized)
A Q4_K_M GGUF of Qwen/Qwen3.5-2B for on-device inference on 6 GB+ iPhones, iPads, and Apple Silicon Macs via llama.cpp or apps that wrap it (e.g. Haplo).
Hosted by jc-builds for the Haplo ecosystem. Quantization by Unsloth. Original weights © Alibaba Cloud, redistributed under the Apache 2.0 License.
TL;DR
A 2B member of Alibaba's Qwen3.5 family with the hybrid Gated DeltaNet + attention architecture. It is the best balance of quality and size for 6 GB iPhones, scoring far above the previous-generation 1.5B models on knowledge and instruction-following benchmarks. It runs in non-thinking mode by default; the upstream card notes the 2B can loop in thinking mode, so keep thinking budgets capped.
Available quantizations
| File | Size | Recommended use |
|---|---|---|
Qwen3.5-2B-Q4_K_M.gguf |
1.28 GB | Default — 6 GB+ devices |
Details
| Parameters | 2B |
| Architecture | qwen35 (hybrid Gated DeltaNet + attention) |
| Quantization | Q4_K_M |
| Chat format | ChatML with an optional <think> block (non-thinking by default) |
| Minimum device | 6 GB RAM (iPhone 14 / 15 class and newer) |
How to use
Haplo (iPhone / iPad / Mac)
The model appears automatically in Haplo's model browser. Download URL:
https://huggingface.co/jc-builds/Qwen3.5-2B-Q4_K_M-GGUF/resolve/main/Qwen3.5-2B-Q4_K_M.gguf
llama.cpp
llama-cli -hf jc-builds/Qwen3.5-2B-Q4_K_M-GGUF:Q4_K_M
License
Apache 2.0. Qwen3.5 by Alibaba Cloud — see the upstream license.