diff --git a/torchtitan/models/llama3/__init__.py b/torchtitan/models/llama3/__init__.py
index 7cfc03b4a..c63bb07c2 100644
--- a/torchtitan/models/llama3/__init__.py
+++ b/torchtitan/models/llama3/__init__.py
@@ -39,6 +39,24 @@
         use_flex_attn=True,
         attn_mask_type="block_causal",
     ),
+    "1B": TransformerModelArgs(
+        dim=2048,
+        n_layers=16,
+        n_heads=32,
+        n_kv_heads=8,
+        ffn_dim_multiplier=1.5,
+        multiple_of=1024,
+        rope_theta=500000,
+    ),
+    "3B": TransformerModelArgs(
+        dim=3072,
+        n_layers=28,
+        n_heads=24,
+        n_kv_heads=8,
+        ffn_dim_multiplier=1.0,
+        multiple_of=1024,
+        rope_theta=500000,
+    ),
     "8B": TransformerModelArgs(
         dim=4096,
         n_layers=32,
diff --git a/torchtitan/models/llama3/train_configs/llama3_1b.toml b/torchtitan/models/llama3/train_configs/llama3_1b.toml
new file mode 100644
index 000000000..116c57762
--- /dev/null
+++ b/torchtitan/models/llama3/train_configs/llama3_1b.toml
@@ -0,0 +1,62 @@
+# torchtitan Config.toml
+# NOTE: this toml config is a preset for 64 A100 GPUs.
+
+[job]
+dump_folder = "./outputs"
+description = "Llama 3 1B training"
+
+[profiling]
+enable_profiling = true
+save_traces_folder = "profile_trace"
+profile_freq = 100
+
+[metrics]
+log_freq = 10
+enable_tensorboard = true
+save_tb_folder = "tb"
+
+[model]
+name = "llama3"
+flavor = "1B"
+tokenizer_path = "./assets/tokenizer/original/tokenizer.model"
+# converters = ["float8"]
+
+[optimizer]
+name = "AdamW"
+lr = 3e-4
+eps = 1e-8
+
+[lr_scheduler]
+warmup_steps = 200  # lr scheduler warm up
+
+[training]
+local_batch_size = 1
+seq_len = 8192
+max_norm = 1.0  # grad norm clipping
+steps = 1000
+compile = false
+dataset = "c4"
+
+[parallelism]
+data_parallel_replicate_degree = 1
+data_parallel_shard_degree = -1
+tensor_parallel_degree = 1
+pipeline_parallel_degree = 1
+context_parallel_degree = 1
+
+[checkpoint]
+enable_checkpoint = false
+folder = "checkpoint"
+interval = 500
+last_save_model_weights_only = false
+export_dtype = "float32"
+async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]
+
+[activation_checkpoint]
+mode = "selective"  # ["none", "selective", "full"]
+selective_ac_option = "op"  # "int" = ac every positive int layer or 'op', ac based on ops policy
+
+[float8]
+enable_fsdp_float8_all_gather = false
+precompute_float8_dynamic_scale_for_fsdp = false
+filter_fqns = ["output"]
diff --git a/torchtitan/models/llama3/train_configs/llama3_3b.toml b/torchtitan/models/llama3/train_configs/llama3_3b.toml
new file mode 100644
index 000000000..3aff16471
--- /dev/null
+++ b/torchtitan/models/llama3/train_configs/llama3_3b.toml
@@ -0,0 +1,62 @@
+# torchtitan Config.toml
+# NOTE: this toml config is a preset for 64 A100 GPUs.
+
+[job]
+dump_folder = "./outputs"
+description = "Llama 3 3B training"
+
+[profiling]
+enable_profiling = true
+save_traces_folder = "profile_trace"
+profile_freq = 100
+
+[metrics]
+log_freq = 10
+enable_tensorboard = true
+save_tb_folder = "tb"
+
+[model]
+name = "llama3"
+flavor = "3B"
+tokenizer_path = "./assets/tokenizer/original/tokenizer.model"
+# converters = ["float8"]
+
+[optimizer]
+name = "AdamW"
+lr = 3e-4
+eps = 1e-8
+
+[lr_scheduler]
+warmup_steps = 200  # lr scheduler warm up
+
+[training]
+local_batch_size = 1
+seq_len = 8192
+max_norm = 1.0  # grad norm clipping
+steps = 1000
+compile = false
+dataset = "c4"
+
+[parallelism]
+data_parallel_replicate_degree = 1
+data_parallel_shard_degree = -1
+tensor_parallel_degree = 1
+pipeline_parallel_degree = 1
+context_parallel_degree = 1
+
+[checkpoint]
+enable_checkpoint = false
+folder = "checkpoint"
+interval = 500
+last_save_model_weights_only = false
+export_dtype = "float32"
+async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]
+
+[activation_checkpoint]
+mode = "selective"  # ["none", "selective", "full"]
+selective_ac_option = "op"  # "int" = ac every positive int layer or 'op', ac based on ops policy
+
+[float8]
+enable_fsdp_float8_all_gather = false
+precompute_float8_dynamic_scale_for_fsdp = false
+filter_fqns = ["output"]