Geode · Qwen3.5-2B for phones
A light model pack for Geode, an Android app that runs language models entirely on the phone: no cloud, no account. Qwen3.5-2B is a small dense model of the same family as Geode's main model (Qwen3.6-35B-A3B): the same hybrid Gated DeltaNet + attention design, the same vocabulary, images understood. It is much faster and needs far less memory, at the cost of knowing and reasoning much less than the 35B.
Packs
| Folder | Weights | Size | Quality (KL vs BF16 / top-1) | For |
|---|---|---|---|---|
Q4/ |
every matrix 4-bit (GPTQ), embedding 8-bit, vision encoder 8-bit, MTP layer 4-bit | 2.0 GB (1.86 GiB) | 0.055 / 90.0% | any 64-bit phone with 4-6 GB of RAM or more; on Snapdragon 8 Gen 2 and newer the NPU reads prompts |
Quality per kind of held-out text (KL divergence from the BF16 model / how often the top token is the same):
| Code | Documents | English | Other languages | Agent session | Tool use | Reasoning | 30 languages |
|---|---|---|---|---|---|---|---|
| 0.023 / 93.6% | 0.038 / 87.5% | 0.042 / 87.0% | 0.034 / 92.5% | 0.158 / 89.9% | 0.055 / 88.9% | 0.035 / 90.3% | 0.058 / 90.6% |
Quantization: GPTQ layer by layer (each layer on what the quantized layers before it produce), calibrated on chat, code, English and Ukrainian text, agent conversations with tool calls, the model's own reasoning and 30 languages.
Speed on a Snapdragon 8 Elite phone (RedMagic 10 Pro): answers ~36 tokens/s, ~54 with the MTP layer's drafts (the same answer, every draft checked); a 2,500-token prompt ~140 tokens/s on the CPU, ~460 with the NPU. Pictures: the image encoder is in the pack; small text in images is read worse than by the 35B.
The packs are Geode's own format (geode.json describes the files). They are not GGUF and do not load in other
runtimes.
Use
- Install Geode (Google Play beta: geode.tifu.xyz).
- Models → pick Qwen3.5-2B in Geode's catalog (or Download from Hugging Face →
Maximilian228/Geode-Qwen3.5-2B).
Requirements: Android 11+, a 64-bit ARM processor with the dot-product instructions (ARMv8.2), and free storage for the pack.
License
Apache-2.0, as the base model: this pack is a quantized derivative of
Qwen/Qwen3.5-2B by the Qwen team (see LICENSE). Answers come from an AI
model and can be wrong.