diff --git a/explorations/asic_quant_combinations.yaml b/explorations/asic_quant_combinations.yaml
new file mode 100644
index 0000000000..ae29206990
--- /dev/null
+++ b/explorations/asic_quant_combinations.yaml
@@ -0,0 +1,110 @@
+# ── asic_quant_combinations.yaml ──────────────────────────────────────────────
+
+named_static_groups:
+  - named_group: "slerp"
+    named_group_settings:
+      mlp_residual_combination: ["slerp"]
+      attn_residual_combination: ["slerp"]
+  - named_group: "add"
+    named_group_settings:
+      mlp_residual_combination: ["add"]
+      attn_residual_combination: ["add"]
+  - named_group: "lerp"
+    named_group_settings:
+      mlp_residual_combination: ["lerp"]
+      attn_residual_combination: ["lerp"]
+  - named_group: "rmsnorm_no_wte"
+    named_group_settings:
+      norm_variant_attn: ["rmsnorm"]
+      norm_variant_output: ["rmsnorm"]
+  - named_group: "rmsnorm_with_wte"
+    named_group_settings:
+      norm_variant_wte: ["rmsnorm"]
+      norm_variant_attn: ["rmsnorm"]
+      norm_variant_output: ["rmsnorm"]
+  - named_group: "offchip_peri_ln"
+    named_group_settings:
+      use_offchip_peri_ln: [true]
+  - named_group: "onchip_peri_ln"
+    named_group_settings:
+      use_peri_ln: [true]
+  - named_group: "no_peri_ln"
+    named_group_settings:
+      use_peri_ln: [false]
+
+common_group:
+  use_edgellm_asic: [true]
+  mlp_variant: ["edgellm_asic_mlp"]
+  attention_variant: ["edgellm_asic_attn"]
+
+  # --- Training & dataset ---
+  dataset: ["minipile"]
+  max_iters: [20000]
+  full_quant_iteration: [10000]
+
+  # --- Model architecture ---
+  n_layer: [8]
+  n_head: [8]
+  n_embd: [512]
+  block_size: [256]
+  batch_size: [64]
+  bias: [false]
+  dtype: ["bfloat16"]
+
+  # --- Quantization settings ---
+  quantization_warmup_iters: [0]
+  quant_scheduler: ["linear"]
+  linear_variant_attn: ["quantized_linear"]
+  linear_variant_mlp: ["quantized_linear"]
+  quantize_linear_method: ["symmetric_quant"]
+  activations_quant_method: ["symmetric_quant"]
+
+  # --- ASIC quantization controls ---
+  quantize_attn_act: [true]
+  quantize_mlp_act: [true]
+  quantize_asic_prenorm: [true]
+  quantize_asic_attn_softmax_denom: [true]
+  quantize_asic_attn_softmax_denom_bits: [16]
+  quantize_asic_attn_softmax_numerator: [true]
+  quantize_asic_attn_softmax_numerator_bits: [8]
+
+  # --- Normalization / regularization ---
+  use_flash_norm: [true]
+  use_pre_ln: [false]
+  dropout: [0.0]
+
+  # --- Optimization ---
+  grad_clip: [1.0]
+  beta1: [0.95]
+  beta2: [0.95]
+  weight_decay: [0.05]
+  learning_rate: [0.00075]
+
+  # --- Sampling / evaluation ---
+  max_sample_tokens: [100]
+  sample_each_eval: [true]
+
+named_variation_groups:
+  - named_group: "residual_combinations"
+    named_group_alternates: ["lerp", "add", "slerp"]
+
+  - named_group: "wte_norms"
+    named_group_alternates: ["rmsnorm_no_wte", "rmsnorm_with_wte"]
+
+  - named_group: "peri_ln_types"
+    named_group_alternates: ["offchip_peri_ln", "onchip_peri_ln", "no_peri_ln"]
+
+parameter_groups:
+  - out_dir: ["asic_quant_combinations"]
+
+    named_group_variations:
+      - "residual_combinations"
+      - "wte_norms"
+      - "peri_ln_types"
+
+    # --- Activation transition ---
+    use_gradual_activation: [true]
+    activation_start: ["gelu"]
+    activation_end: ["relu"]
+    activation_transition_start_iter: [0]
+    activation_transition_end_iter: [20000]