Add LFM3-1.7B MLX build (int4, Apple GPU) and update the notice

.gitattributes

@@ -52,3 +52,4 @@ sam3/coreml/sam3_vision2_672_coreml.pte filter=lfs diff=lfs merge=lfs -text
 sam3/coreml/sam3_vision3_672_coreml.pte filter=lfs diff=lfs merge=lfs -text
 lfm3/tokenizer.json filter=lfs diff=lfs merge=lfs -text
 lfm3/xnnpack/lfm3_1_7b_xnnpack_8da4w.pte filter=lfs diff=lfs merge=lfs -text
+lfm3/mlx/lfm3_1_7b_mlx_4w.pte filter=lfs diff=lfs merge=lfs -text

lfm3/NOTICE.md

@@ -1,5 +1,6 @@
 These files are modified derivatives of LiquidAI/LFM3-1.7B (LFM Open License v1.0, see LICENSE in this folder):
 
-- `xnnpack/lfm3_1_7b_xnnpack_8da4w.pte`: the weights re-implemented for ExecuTorch export and quantized (linear layers int8 dynamic activations / int4 weights, group 128; embedding int8).
+- `xnnpack/lfm3_1_7b_xnnpack_8da4w.pte`: the weights re-implemented for ExecuTorch export and quantized (linear layers int8 dynamic activations / int4 weights, group 128; embedding int8). CPU, iOS and Android.
+- `mlx/lfm3_1_7b_mlx_4w.pte`: same export for the MLX backend (Apple GPU), bf16 compute with int4 weight-only linears and embedding, group 64.
 - `tokenizer_config.json`: Liquid's tokenizer config with the contents of `chat_template.jinja` embedded as `chat_template`, unchanged otherwise.
 - `tokenizer.json`: unchanged.

lfm3/mlx/lfm3_1_7b_mlx_4w.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96df77d94a70194e619c4aa0e871662f724bf9836e85bfeec9d2d737550c6dcc
+size 1066865408