Geode · Qwen3.8-Flash-Next for phones
A model pack for Geode, an Android app that runs language models entirely on the phone: no cloud, no account. Qwen3.8-Flash-Next is the Qwen team's 180 B-parameter mixture of experts: 48 layers (Gated DeltaNet and attention with a sparse indexer), four hyper-connection streams, 512 experts of which 10 answer each token, and a 51 B n-gram embedding table read a few rows a token. It is by far the largest model Geode runs and it is slow on a phone: an extra for the biggest phones, there because it can be.
Pack
| Folder | Weights | Size | For |
|---|---|---|---|
Q4/ |
routed experts and attention / DeltaNet / shared-expert matrices 4-bit (Geode's q4: an FP8 scale per 16), quantized by GPTQ straight from Qwen's BF16 checkpoint; lm_head and the embedding 8-bit from BF16; routers, hyper-connections, the indexer, the n-gram projections bf16; the n-gram table in one of two variants (below); the vision encoder and the MTP draft layer | 125 GB with the FP8 table, 103 GB with the q4 table | 64-bit phones with 12 GB of RAM or more and that much free storage |
The n-gram table comes in two variants; the app downloads one:
| Variant | Files | Size | |
|---|---|---|---|
| FP8 (default) | ngram-f8-*.bin |
51.2 GB | Qwen's own FP8 bytes, bit for bit |
| q4 | ngram-q4-*.bin |
28.8 GB | saves 22 GB; the answers are slightly less accurate |
Only a few rows of the table and ten experts a layer are read for each token, so the pack runs from flash with a small part of it in RAM; reading those experts from storage is what sets the speed.
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 → Download from Hugging Face →
Maximilian228/Geode-Qwen3.8-Flash-Next, choose the n-gram table.
Requirements: Android 11+, a 64-bit ARM processor with the dot-product instructions (ARMv8.2), 12 GB of RAM or more, and free storage for the pack.
License and attribution
The weights are the Qwen team's Qwen3.8-Flash-Next, under the
Qwen Community License 1.0 (LICENSE, copied unchanged); this pack is a derivative work under the same license, and
its terms (the notice; the separate license Qwen requires for commercial Model-as-a-Service or AI Work Assistant use)
apply to it. Changes made here: the model converted into Geode's file layout and quantized - the routed experts and
the attention / DeltaNet / shared-expert matrices quantized to 4 bits by GPTQ from the BF16 checkpoint, lm_head and
the embedding rounded to 8 bits, the n-gram table kept as FP8 or rounded to 4 bits. Answers come from an
AI model and can be wrong.
Model tree for Maximilian228/Geode-Qwen3.8-Flash-Next
Base model
Qwen/Qwen3.8-Flash-Next