Downloads · 30 days
68
11% of all-time downloads
Gemma-4-E2B-it-heretic-ara-MNN
darkmaniac7/Gemma-4-E2B-it-heretic-ara-MNN
Gemma-4-E2B-it-heretic-ara-MNN is a any-to-any model from darkmaniac7. Use it for the any-to-any task on the model card, and read the license before you ship it in a product. It is set up for mnn. The card lists the license as apache-2.0.
p-e-w's Heretic ARA abliteration of Google's Gemma 4 E2B it, as a 4-bit MNN bundle for chatting on your phone with TokForge.
All-time downloads
605
Public
Repo size
6.8 GB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.weight2.2 GB · 60%
At a glance
- Task
- Any-to-Any
- Library
- mnn
- License
- apache-2.0
- Access
- Public
- Created
- Apr 12, 2026
- Updated
- Nov 11, 2026
- SHA
- 4789059f
Base models
Model card
From the Hugging Face model README
Gemma 4 E2B it heretic ara (MNN)
p-e-w's Heretic ARA abliteration of Google's Gemma 4 E2B it, as a 4-bit MNN bundle for chatting on your phone with TokForge.
Published by TokForge for on-device inference.
| What it is | Vision chat model (text + images; abliterated) |
| Format | MNN bundle, 4-bit HQQ weights (block 64), with vision and audio encoders |
| Runs on | Android phones (CPU); also listed in TokForge for iOS |
| Phone class | 8 GB RAM or more (TokForge for iOS catalog; the Android app does not list it today) |
| Download | 3.73 GB (12 files) |
| Base model | p-e-w/gemma-4-E2B-it-heretic-ara |
| 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, or join the iOS TestFlight.
- TokForge for iOS lists it as Gemma 4 E2B Heretic Ara (MNN) for devices with 8 GB RAM or more. The Android app does not list it under Recommended today; on Android, open Models > Browse, search for
Gemma-4-E2B-it-heretic-ara-MNNand download it. - Load it and start a chat. The model runs on the phone; once it is downloaded, chatting needs no connection.
Collections: TokForge Gemma Models (MNN), TokForge Vision Models.
Files
| File | Size | What it is |
|---|---|---|
audio.mnn | 1.43 MB | Audio encoder graph |
audio.mnn.weight | 590 MB | Audio encoder weights |
config.json | 694 B | MNN runtime settings |
export_args.json | 1.20 KB | Export settings used for the conversion |
llm.mnn | 2.13 MB | Model graph |
llm.mnn.json | 4.61 MB | Model graph as JSON (reference copy) |
llm.mnn.weight | 1.42 GB | Model weights |
llm_config.json | 1.42 KB | Model settings and chat template |
ple_embeddings_int4.bin | 1.47 GB | Per-layer embeddings (int4) |
tokenizer.mtok | 10.1 MB | Tokenizer |
visual.mnn | 1.06 MB | Vision encoder graph |
visual.mnn.weight | 226 MB | Vision encoder weights |
About the model
The Heretic ARA abliterated variant of google/gemma-4-E2B-it, made by p-e-w with Heretic v1.2.0 (Arbitrary-Rank Ablation, row-norm preservation). Abliteration edits the weights so the model refuses fewer requests; you are responsible for how you use it.
| Architecture | Gemma 4 (35 layers, hidden size 1536, sliding-window and full attention) |
| Quantization | 4-bit HQQ weights, block size 64; per-layer embeddings 4-bit in ple_embeddings_int4.bin |
| Encoders | Vision (visual.mnn) and audio (audio.mnn) |
| Chat template | Gemma 4 turn format, in llm_config.json |
Vision and audio
On Android, TokForge reads photos you attach to a chat with this bundle's own vision encoder. Gemma 4's audio encoder is included as well.
Settings TokForge uses
TokForge applies these by itself on Android; you do not need to set them.
| Setting | Value |
|---|---|
| Backend | CPU |
| Precision | Normal or higher (TokForge raises Low to Normal for Gemma 4) |
| KV cache | Default (no quantized KV cache for Gemma 4) |
| Thinking | Off |
| Speculative decoding | Off |
How it was made
- Exported on 2026-04-25 with MNN's llmexport (TokForge's MNN build).
- Same file layout as taobao-mnn/gemma-4-E2B-it-MNN, the layout TokForge downloads for Gemma 4.
llm.mnn+llm.mnn.weight: 4-bit HQQ weights, block size 64.ple_embeddings_int4.bin: per-layer embeddings, 4-bit.tokenizer.mtok: MNN's binary tokenizer.visual.mnn+audio.mnn: the multimodal encoders.llm_config.json:model_type: gemma4, withple_embed_file+ple_quant.
Limitations
- At Low precision the model can repeat the prompt back instead of answering; it gives coherent replies at Normal or High precision, which is what TokForge uses.
- This is a runtime bundle, not a standard Transformers training checkpoint.
- Quantization (Q4) may slightly reduce quality compared to the full-precision original.
License
Apache-2.0, as Google releases Gemma 4 (license).
Credits
- gemma-4-E2B-it-heretic-ara by p-e-w (Hugging Face). Used under the Apache-2.0 license. Our MNN files were converted from this model.
- Gemma 4 E2B IT by Google (Hugging Face). Used under the Apache-2.0 license.
- Abliterated with Heretic by p-e-w.
- Converted with llmexport from MNN by Alibaba (Apache-2.0).
- Task
- Any-to-Any
- Library
- mnn
- License
- apache-2.0
- Languages
- en
- Created
- Apr 12, 2026
- Updated
- Nov 11, 2026