TokForge NPU packs (index)
The list of NPU prompt packs that TokForge can offer on Snapdragon phones, kept as one small JSON file that the app reads.
Published by TokForge for on-device inference.
| What it is | Index file |
| Format | One JSON file, npu_pack_index.json (schema tokforge-npu-index/1) |
| Runs on | Read by TokForge 1.3.9.1 and later on Android; the packs it lists run on Snapdragon phones (Hexagon NPU) |
| Phone class | Not a model; the packs it lists need a Snapdragon 8 Elite Gen 5, Snapdragon 8 Elite or Snapdragon 8 Gen 3 |
| Download | 7.17 KB (1 file) |
| Base model | None (a list of links and hashes) |
| License | Apache-2.0 |
TokForge
- Website: https://tokforge.ai
- Discord: https://discord.gg/Acv3CBtfVm
- Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
- iOS TestFlight: https://testflight.apple.com/join/jnufjzRr
Runs on-device in the TokForge app.
Use it in TokForge
- Install TokForge from Google Play.
- You never download this file by hand. When you open Models > Browse and the copy on the phone is older than 12 hours, TokForge fetches
npu_pack_index.jsonfrommainand lists the packs for your chip under NPU packs for your phone. - Tap Get next to a pack for a model you have downloaded. TokForge downloads the pack from its own repo at the commit named in this file and checks it before it uses the NPU.
Collections: TokForge NPU Packs.
Files
| File | Size | What it is |
|---|---|---|
npu_pack_index.json |
7.17 KB | The pack index TokForge reads (schema tokforge-npu-index/1, one entry per pack) |
How TokForge uses it
This repo holds one small file, npu_pack_index.json: the list of NPU prompt packs that TokForge can offer on Snapdragon
phones. An NPU prompt pack lets the phone's NPU read long prompts while the CPU writes every reply with the model you
already downloaded. Each pack lives in its own repo; this file says which packs exist and where.
The TokForge app reads this file from main for its NPU packs for your phone list in Models > Browse, keeps a copy on the phone and checks
for a newer one now and then. The file never decides on its own what runs: every entry points at a pack repo at one
fixed commit, and the app checks the pack's MANIFEST.json against the sha256 written here, then every file of the pack
by sha256 and size, and the pairing with your download, before it uses the pack. A later change to this file can add a
pack; it cannot change the files of a pack your phone already checked.
Packs in the index
| Model | Pack repo (pinned commit) | Download per chip | NPU memory per chip | Memory figure |
|---|---|---|---|---|
| Qwen3.5 4B | darkmaniac7/Qwen3.5-4B-NPU-catalogpair at 43e528cc1ce4 |
V81 3.67 GB, V79 3.65 GB, V75 3.65 GB | V81 6.31 GB, V79 5.84 GB, V75 6.26 GB | measured on phones for V81, V79 and V75 |
| Qwen3.5 4B Abliterated | darkmaniac7/Qwen3.5-4B-uncensored-NPU-catalogpair at d384c68e4122 |
V81 3.67 GB, V75 3.65 GB | V81 5.93 GB, V75 5.87 GB | measured on phones for V81 and V75 |
| Qwen3.5 2B | darkmaniac7/Qwen3.5-2B-NPU-catalogpair at 5752109061a2 |
V81 1.41 GB, V75 1.40 GB | V81 2.93 GB, V75 2.90 GB | measured on phones for V81 and V75 |
| Qwen3.5 4B Claude 4.6 Opus Abliterated | darkmaniac7/Qwen3.5-4B-Claude-4.6-Opus-abliterated-NPU-catalogpair at e3378b577741 |
V81 3.67 GB, V75 3.65 GB | V81 5.90 GB, V75 5.85 GB | measured on phones for V81 and V75 |
| Qwen3.5 0.8B | darkmaniac7/Qwen3.5-0.8B-NPU-catalogpair at 56aac1e38f03 |
V81 0.53 GB, V75 0.52 GB | V81 2.17 GB, V75 2.10 GB | measured on phones for V81 and V75 |
| Qwen3.5 9B (QAIRT 2.51) | darkmaniac7/Qwen3.5-9B-NPU-catalogpair at 9daa5ec6f672 |
V81 3.97 GB | V81 6.63 GB | measured on phones for V81 |
| Qwen3.5 4B W4 (QAIRT 2.51) | darkmaniac7/Qwen3.5-4B-NPU-catalogpair-qairt251 at e22fbfefdcd8 |
V81 2.09 GB, V79 2.08 GB, V75 2.08 GB | V81 4.63 GB, V79 4.34 GB, V75 4.63 GB | measured on phones for V81 and V79; V75 provisional (the V81 figure), not tested on a phone yet |
V81 is the Snapdragon 8 Elite Gen 5 (SM8850), V79 the Snapdragon 8 Elite (SM8750), V75 the Snapdragon 8 Gen 3 (SM8650). A chip
that has a download for a pack but is not in its measured_arches has not been tested on a phone by us yet; TokForge says so
and checks the pack on the phone before it uses it, as on every chip. The memory figure is what the NPU
part takes while the model is loaded, on top of the model itself; TokForge turns the NPU on only when the phone has room
for both. Where a figure is estimated, the app measures the real one when it first loads the pack on your phone and
uses that from then on.
Packs marked QAIRT 2.51 were prepared with the QAIRT 2.51 NPU runtime, which TokForge carries from version 1.3.9.6 (the app loads that runtime for them after one restart); older versions of the app skip those entries. The 9B pack is V81 only and needs a 24 GB phone. The 4B W4 pack is the Qwen3.5 4B at its exact catalog weights in a smaller download with less NPU memory; the 4B pack above stays for the runtime every version of the app carries.
The Qwen3 packs (Qwen3 0.6B, Qwen3 1.7B and Qwen3 4B Abliterated) are not in this file: TokForge carries their pins itself and downloads them from darkmaniac7/TokForge-LLM-QNN-NPU.
What each entry holds
Schema tokforge-npu-index/1: {"schema": "tokforge-npu-index/1", "packs": [...]}, one object per pack:
model_id: the TokForge model list row the pack belongs to.name: the pack folder inside its repo (<name>/MANIFEST.json,<name>/<chip>/...).family: the model family (qwen3_5orqwen3); a pack built for a newer NPU runtime adds its runtime line (qwen3_5/qairt-2.51), and an app that does not know that token skips the entry.qnn_sdk_version: for aqairt-entry, the exact QAIRT runtime version the pack was prepared with.pack_repo,pack_revision: the pack repo and the 40-character commit the app downloads from.manifest_sha256: the sha256 of the pack's MANIFEST.json at that commit.arch_bytes: the download size per chip.resident_bytes: the NPU part's memory per chip;measured_arches: the chips where that figure was measured on a phone.pairs_with: the download the pack works with (repo, commit, llm.mnn sha256, llm.mnn.weight size and sha256, and embeddings_bf16.bin when that model keeps its embeddings in a separate file).
Every version of this file is checked against the Hub before it is pushed: each MANIFEST.json hash, its pairing and its per-chip sizes.
License
Apache License 2.0 (https://www.apache.org/licenses/LICENSE-2.0). The index is a list of links and hashes; each pack repo's card names the original models, their authors and their licenses.
Credits
- The packs listed here are built from Qwen3.5 models by the Qwen team (Alibaba Cloud) (Hugging Face) and from derived models by huihui-ai and Jackrong. Each pack repo's card credits its sources and their licences.