Card: blockwise-128 + cache2048 specs (iPhone ~27 tok/s, GSM8K 77%)

README.md

@@ -25,12 +25,12 @@ Text-only conversion (the Falcon3 decoder; no vision/audio towers).
 
 | | |
 |---|---|
-| **File** | `model.litertlm` (~1.86 GB) |
-| **Quantization** | int4 weights — **blockwise (block 32)**, symmetric; embeddings INT8 |
+| **File** | `model.litertlm` (~1.74 GB) |
+| **Quantization** | int4 weights — **blockwise (block 128)**, symmetric; embeddings INT8 |
 | **Compute** | integer |
-| **Context (KV cache)** | 4096 |
+| **Context (KV cache)** | 2048 |
 | **Base model** | tiiuae/Falcon3-3B-Instruct |
-| **Decode speed** | ~101 tok/s (Apple Silicon, CPU, LiteRT-LM, greedy) |
+| **Decode speed** | ~27 tok/s (iPhone 17 Pro, Metal GPU) · ~89 tok/s (Mac M4 Max, LiteRT-LM, greedy) |
 
 ## Usage
 
@@ -58,7 +58,7 @@ known-good 4-bit control:
 |---|---|
 | bf16 (reference) | 75% |
 | MLX 4-bit (control) | 76% |
-| **This model — LiteRT int4** | **79%** |
+| **This model — LiteRT int4** | **77%** |
 
 LiteRT int4 is fully at parity — it matches or slightly exceeds both the 4-bit
 control and bf16 here (the small spread is sampling noise at n=100). This is a
@@ -68,8 +68,8 @@ direct-answering instruct model (no `<think>` block) and terminates cleanly at
 ## Conversion
 
 Converted with [`litert-torch`](https://github.com/google-ai-edge/litert) using a
-**blockwise int4** recipe (INT4 weights, block size 32, symmetric) with embeddings
-kept at INT8, KV cache 4096, and Falcon3's native chat template. Falcon3-3B is a
+**blockwise int4** recipe (INT4 weights, block size 128, symmetric) with embeddings
+kept at INT8, KV cache 2048, and Falcon3's native chat template. Falcon3-3B is a
 standard `LlamaForCausalLM` architecture, so it rides the existing converter and
 runtime with no custom code. Blockwise (not channelwise) int4 is what preserves
 reasoning accuracy.