Upload README.md with huggingface_hub

README.md

@@ -9,42 +9,33 @@ tags:
   - gated-deltanet
 ---
 
-# Qwen 3.5 2B — TFLite
+# Qwen 3.5 2B — TFLite (.tflite)
 
-Qwen 3.5 2B converted to **TFLite** format (.tflite) for on-device inference.
+Qwen 3.5 2B as raw TFLite format for on-device inference with the TFLite Interpreter API.
 
-**First working conversion of Qwen 3.5's hybrid GatedDeltaNet + full attention architecture to TFLite.**
+> **For LiteRT-LM Engine usage, use the bundled version instead: [paulsp94/Qwen3.5-2B-LiteRT-LM](https://huggingface.co/paulsp94/Qwen3.5-2B-LiteRT-LM)**
 
-## Format
+## What's this
 
-This is a **raw .tflite file**, not a .litertlm bundle. To use with LiteRT-LM's `Engine`, you'll need to bundle it with a tokenizer into .litertlm format. It can also be used directly with the TFLite Interpreter API.
+Raw `.tflite` model file — use this if you're building your own inference pipeline with the TFLite Interpreter API directly. If you want the ready-to-use LiteRT-LM bundle with tokenizer included, use the [LiteRT-LM version](https://huggingface.co/paulsp94/Qwen3.5-2B-LiteRT-LM) instead.
 
-## Model Details
+## Architecture
 
 | | |
 |---|---|
 | **Base model** | [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) |
-| **Architecture** | 24 layers — 18x GatedDeltaNet linear attention + 6x standard GQA |
+| **Layers** | 24 total: 18× GatedDeltaNet linear + 6× GQA full attention |
 | **Quantization** | int8 dynamic |
 | **Format** | TFLite (.tflite) |
 | **Size** | ~1.9 GB |
 
-## Architecture Notes
-
-Qwen 3.5 uses a hybrid architecture that no other converter previously supported:
-
-- **Linear attention (GatedDeltaNet):** Recurrent state-space model with A_log decay, conv1d, output gating — 18 layers
-- **Full attention (every 4th layer):** Grouped query attention with asymmetric Q/KV head dims (Q=512 with partial rotary, KV=256) — 6 layers
-
-Both layer types implemented using standard TFLite ops (matmul, element-wise, conv1d) — no custom kernels.
-
 ## Files
 
-- `qwen35_2b.tflite` — The converted model (1.9 GB)
-- `tokenizer.json` — BPE tokenizer (from original Qwen 3.5 2B)
+- `qwen35_2b.tflite` — The converted model
+- `tokenizer.json` — BPE tokenizer (you'll need to handle tokenization yourself)
 - `tokenizer_config.json` — Tokenizer configuration
 - `config.json` — Original model config
 
 ## Conversion
 
-Converted using custom litert-torch model authoring with GatedDeltaNet linear attention implementation. Source code: [allot/tools/model-export](https://github.com/paulsp94/allot/tree/main/tools/model-export)
+Source: [allot/tools/model-export](https://github.com/paulsp94/allot/tree/main/tools/model-export)