diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 83f9118da..087376502 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -287,6 +287,26 @@ typedef struct {
 } block_tq2_0;
 static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0 block size/padding");
 
+//
+// Dual-plane ternary quantization (DT3)
+//
+// w_i = d[0]*t0_i + d[1]*t1_i, with t in {-1, 0, +1}. Two ternary planes buy
+// back most of the quality a single one loses, at 3.5 bpw instead of 1.6875.
+//
+// Each plane is packed exactly like tq1_0 with every constant halved, since the
+// block is 128 elements instead of QK_K: qs 48 -> 24 bytes, qh 4 -> 2 bytes,
+// and the two qs passes go over 16 and then 8 bytes instead of 32 and 16.
+//
+#define QK_DT3 128
+
+// 3.5 bpw
+typedef struct {
+    uint8_t qs[2][(QK_DT3 - 4 * QK_DT3 / 64) / 5]; // 2 planes, 5 elements per byte (3^5 = 243 < 256)
+    uint8_t qh[2][QK_DT3 / 64];                    // 2 planes, 4 elements per byte
+    ggml_half d[2];                                // one scale per plane
+} block_dt3;
+static_assert(sizeof(block_dt3) == 2 * (sizeof(ggml_half) + QK_DT3 / 64 + (QK_DT3 - 4 * QK_DT3 / 64) / 5), "wrong dt3 block size/padding");
+
 //
 // Super-block quantization structures
 //