diff --git a/changes/render3d-bc-textures.md b/changes/render3d-bc-textures.md
new file mode 100644
index 00000000..15c85544
--- /dev/null
+++ b/changes/render3d-bc-textures.md
@@ -0,0 +1,11 @@
+bump: minor
+type: feature
+**A texture with a `.dds` beside its `.png` goes to the GPU BC-compressed, its mips included.**
+`tex_load` (and everything built on it: glTF materials, the terrain's materials, ludic.ui's
+pictures) looks for `<name>.dds` when the device has `textureCompressionBC` - every desktop GPU,
+Apple silicon through MoltenVK too - and uploads its blocks as they are: BC7 (colour in sRGB,
+or data read as it is), BC5 or BC4, a quarter of RGBA8 or less in video memory, no decoding and
+no mipmapping on the load. Only a DX10 `.dds` carrying the whole chain to 1x1 is taken; anything
+else is reported and the `.png` is used, as it is for a cut-out atlas (padded from the `.png`).
+`gpu_tex_compressed` and `R3D_BC7` / `R3D_BC7_SRGB` / `R3D_BC5` / `R3D_BC4` are the backend's
+entry; `examples/rendering/bc.ludic` is the check, and ludic.lab's plate carries its `.dds`.
diff --git a/examples/rendering/bc.ludic b/examples/rendering/bc.ludic
new file mode 100644
index 00000000..276e91d8
--- /dev/null
+++ b/examples/rendering/bc.ludic
@@ -0,0 +1,59 @@
+# bc.ludic - a texture whose .png has a .dds beside it arrives BC7-compressed with its whole mip
+# chain (plan 23.3 of maroon-lake), and one without a .dds as the .png it always was. Prints BC OK.
+#
+#   bin/ludic build examples/rendering/bc.ludic --headless && ./build/bc_headless
+program Bc {
+  numbers float
+  import "ludic.render3d/r3d.ludic"
+
+  property Marker { on: int = 1 }
+  model Anchor { Marker }
+
+  function scene_draw(render3d_st: mut Render3dState) -> void { }
+  function scene_draw_casters(render3d_st: mut Render3dState, light_vp: floats) -> void { }
+  function stream_fill(s: Stream, cx: int, cz: int, band: int) -> void { }
+
+  function check(what: string, cond: bool) -> bool {
+    if not cond { print(`bc: FAILED - {what}`) }
+    return cond
+  }
+  # the internal format render3d recorded for a texture
+  function fmt_of(render3d_st: mut Render3dState, tex: int) -> int {
+    let o = gpu_tx_at(render3d_st, tex)
+    if o < 0 { return 0 }
+    return render3d_st.gpu_tx[o + 4]
+  }
+
+  handler Boot(render3d_st: mut Render3dState) phase Start {
+    spawn Anchor {}
+    r3d_on_draw(render3d_st, fn scene_draw)
+    r3d_on_casters(render3d_st, fn scene_draw_casters)
+    r3d_on_stream_fill(render3d_st, fn stream_fill)
+    r3d_plate_mode(render3d_st, true)
+    let dir = "packages/ludic.lab/plate"
+    render3d_st.r3d_sky_path = `{dir}/sky.hdr`
+    if not r3d_init(render3d_st, 320, 180, "Bc") {
+      quit()
+      return
+    }
+    var ok = true
+    ok = check("this device reads BC textures", render3d_st.gvk_has_bc) and ok
+    let t = tex_load(render3d_st, `{dir}/plate_diff.png`, true)
+    ok = check("the plate's colour loads", t != 0) and ok
+    ok = check("from its .dds, as BC7 in sRGB", fmt_of(render3d_st, t) == R3D_BC7_SRGB) and ok
+    ok = check("with every level down to 1x1", render3d_st.gvk_tex_levels[t] == gvk_mip_levels(tex_width(render3d_st, t), tex_height(render3d_st, t))) and ok
+    let a = tex_load(render3d_st, `{dir}/plate_arm.png`, false)
+    ok = check("a data map's .dds is BC7 read as it is (not sRGB)", fmt_of(render3d_st, a) == R3D_BC7) and ok
+    let s = tex_load(render3d_st, `{dir}/probe_arm.png`, false)
+    ok = check("the second one loads too", s != 0) and ok
+    # a picture with no .dds: the .png, uncompressed, as before
+    let p = tex_load_ex(render3d_st, `{dir}/plate_diff.png`, true, 1)
+    ok = check("a cut-out atlas (edge-padded from the .png) is not the .dds", p != 0 and fmt_of(render3d_st, p) != R3D_BC7_SRGB) and ok
+    if ok { print("BC OK") } else { print("BC FAILED") }
+    quit()
+  }
+
+  handler Present(render3d_st: mut Render3dState) phase Render {
+    r3d_present(render3d_st)
+  }
+}
diff --git a/packages/ludic.lab/plate/plate_arm.dds b/packages/ludic.lab/plate/plate_arm.dds
new file mode 100644
index 00000000..025b7efa
Binary files /dev/null and b/packages/ludic.lab/plate/plate_arm.dds differ
diff --git a/packages/ludic.lab/plate/plate_diff.dds b/packages/ludic.lab/plate/plate_diff.dds
new file mode 100644
index 00000000..c22d5e80
Binary files /dev/null and b/packages/ludic.lab/plate/plate_diff.dds differ
diff --git a/packages/ludic.lab/plate/probe_arm.dds b/packages/ludic.lab/plate/probe_arm.dds
new file mode 100644
index 00000000..bb0814e4
Binary files /dev/null and b/packages/ludic.lab/plate/probe_arm.dds differ
diff --git a/packages/ludic.lab/plate/probe_ball.dds b/packages/ludic.lab/plate/probe_ball.dds
new file mode 100644
index 00000000..00978249
Binary files /dev/null and b/packages/ludic.lab/plate/probe_ball.dds differ
diff --git a/packages/ludic.lab/plate/probe_box.dds b/packages/ludic.lab/plate/probe_box.dds
new file mode 100644
index 00000000..3faf3939
Binary files /dev/null and b/packages/ludic.lab/plate/probe_box.dds differ
diff --git a/packages/ludic.render3d/env.ludic b/packages/ludic.render3d/env.ludic
index 6ddb84aa..bbeedfdb 100644
--- a/packages/ludic.render3d/env.ludic
+++ b/packages/ludic.render3d/env.ludic
@@ -200,6 +200,7 @@ export state Render3dState {
   gvk_labels: bool = false          # R3D_VK_LABELS: each profiled pass is a debug label (Metal System Trace)
   gvk_inflight: int = -1
   gvk_nopool: bool = false
+  gvk_has_bc: bool = false          # textureCompressionBC: .dds textures are uploaded compressed
   # R3D_VK_PROF: Vulkan objects made and destroyed, reported every 120 frames
   gvk_mk_img: int = 0
   gvk_mk_view: int = 0
diff --git a/packages/ludic.render3d/gpu.ludic b/packages/ludic.render3d/gpu.ludic
index 9def4f8e..df2cd0a3 100644
--- a/packages/ludic.render3d/gpu.ludic
+++ b/packages/ludic.render3d/gpu.ludic
@@ -571,6 +571,18 @@ function gpu_tex_paramf(render3d_st: mut Render3dState, kind: int, pname: int, v
 }
 # the border colour clamp-to-border reads (four fixed values in `rgba`)
 function gpu_tex_border(render3d_st: Render3dState, kind: int, rgba: pointer) -> void { }
+# a block-compressed texture with its whole mip chain, into the bound 2D texture: levels as they
+# are in the file (level l at offs[l]), no decoding on the CPU and no mips made here
+function gpu_tex_compressed(render3d_st: mut Render3dState, ifmt: int, w: int, h: int, levels: int, data: pointer, offs: words) -> bool {
+  gvk_flush(render3d_st)
+  let tex = render3d_st.gpu_bound_2d
+  if not gvk_tex_storage(render3d_st, tex, false, ifmt, w, h, 1, true) { return false }
+  if render3d_st.gvk_tex_levels[tex] != levels { print(`r3d: a {w}x{h} compressed texture brought {levels} levels, not {render3d_st.gvk_tex_levels[tex]}`);  return false }
+  if not gvk_tex_upload_blocks(render3d_st, tex, w, h, levels, data, offs) { return false }
+  let o = gpu_tx_at(render3d_st, tex)
+  if o >= 0 { render3d_st.gpu_tx[o] = GPU_TEX2D;  render3d_st.gpu_tx[o + 1] = w;  render3d_st.gpu_tx[o + 2] = h;  render3d_st.gpu_tx[o + 3] = 1;  render3d_st.gpu_tx[o + 4] = ifmt;  render3d_st.gpu_tx[o + 10] = 1 }
+  return true
+}
 function gpu_tex_mips(render3d_st: mut Render3dState, kind: int) -> void {
   let mt = gpu_bound(render3d_st, kind)
   let mo = gpu_tx_at(render3d_st, mt)
diff --git a/packages/ludic.render3d/gpu_vk.ludic b/packages/ludic.render3d/gpu_vk.ludic
index ca1f38f6..ecbd280c 100644
--- a/packages/ludic.render3d/gpu_vk.ludic
+++ b/packages/ludic.render3d/gpu_vk.ludic
@@ -174,6 +174,10 @@ function gvk_init(render3d_st: mut Render3dState) -> bool {
   # how many of them there are. Asked for where the device has them; the renderer checks
   # gvk_has_mdi / gvk_has_dic before it takes that path.
   render3d_st.gvk_has_mdi = Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_multiDrawIndirect) == 1 and Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_drawIndirectFirstInstance) == 1
+  # BC-compressed textures (BC4/5/7): every desktop GPU, Apple silicon through MoltenVK too; a
+  # texture with a .dds beside its .png is uploaded compressed when this is on (texture.ludic)
+  render3d_st.gvk_has_bc = Vk.get_i32(f2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_textureCompressionBC) == 1
+  if render3d_st.gvk_has_bc { Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_textureCompressionBC, 1) }
   if render3d_st.gvk_has_mdi {
     Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_multiDrawIndirect, 1)
     Vk.put_i32(want2, VkPhysicalDeviceFeatures2_features + VkPhysicalDeviceFeatures_drawIndirectFirstInstance, 1)
diff --git a/packages/ludic.render3d/gpu_vk_bc.ludic b/packages/ludic.render3d/gpu_vk_bc.ludic
new file mode 100644
index 00000000..7e679896
--- /dev/null
+++ b/packages/ludic.render3d/gpu_vk_bc.ludic
@@ -0,0 +1,31 @@
+# gpu_vk_bc.ludic - a block-compressed texture's levels uploaded as they are: the GPU decodes BC
+# blocks itself, so the bytes in the file are the bytes in video memory, a quarter of RGBA8's.
+
+# every level of `tex` (made by gvk_tex_storage with its whole chain) from `data`, level l's
+# blocks at offs[l] .. offs[l + 1]
+function gvk_tex_upload_blocks(render3d_st: mut Render3dState, tex: int, w: int, h: int, levels: int, data: pointer, offs: words) -> bool {
+  let n = offs[levels]
+  let dst = gvk_staging(render3d_st, n, VK_BUFFER_USAGE_TRANSFER_SRC_BIT)
+  if dst == null { print("r3d: vulkan: no staging buffer for a compressed upload");  return false }
+  mem_copy(dst, data, n)
+  let image = render3d_st.gvk_tex_image[tex]
+  let cb = gvk_once_begin(render3d_st)
+  gvk_barrier(render3d_st, cb, image, false, 0, levels, 1, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL)
+  for l in 0 .. levels {
+    let bic = gvk_tmp(render3d_st, VkBufferImageCopy_sizeof)
+    Vk.zero(bic, VkBufferImageCopy_sizeof)
+    let off: long = offs[l]
+    Vk.put_i64(bic, VkBufferImageCopy_bufferOffset, off)
+    Vk.put_i32(bic, VkBufferImageCopy_imageSubresource + VkImageSubresourceLayers_aspectMask, VK_IMAGE_ASPECT_COLOR_BIT)
+    Vk.put_i32(bic, VkBufferImageCopy_imageSubresource + VkImageSubresourceLayers_mipLevel, l)
+    Vk.put_i32(bic, VkBufferImageCopy_imageSubresource + VkImageSubresourceLayers_layerCount, 1)
+    Vk.put_i32(bic, VkBufferImageCopy_imageExtent + VkExtent3D_width, max(w >> l, 1))
+    Vk.put_i32(bic, VkBufferImageCopy_imageExtent + VkExtent3D_height, max(h >> l, 1))
+    Vk.put_i32(bic, VkBufferImageCopy_imageExtent + VkExtent3D_depth, 1)
+    Vk.cmd_copy_buffer_to_image(cb, render3d_st.gvk_st_buf, image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, bic)
+  }
+  gvk_barrier(render3d_st, cb, image, false, 0, levels, 1, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL)
+  let ok = gvk_once_end(render3d_st, cb)
+  gvk_staging_free(render3d_st)
+  return ok
+}
diff --git a/packages/ludic.render3d/gpu_vk_res.ludic b/packages/ludic.render3d/gpu_vk_res.ludic
index b8357821..5a8ad74c 100644
--- a/packages/ludic.render3d/gpu_vk_res.ludic
+++ b/packages/ludic.render3d/gpu_vk_res.ludic
@@ -9,7 +9,18 @@
 # The renderer names formats the way OpenGL does (they are gpu.ludic's vocabulary); these turn
 # one into the Vulkan image format that holds it. Three-channel formats have no widely
 # supported Vulkan image format, so they are stored with four and expanded on upload.
+# block-compressed formats, named here (the GL header generated for render3d carries no BPTC):
+# the values are GL's, so they sit in the same vocabulary as every other internal format
+const R3D_BC7: int = 0x8E8C
+const R3D_BC7_SRGB: int = 0x8E8D
+const R3D_BC5: int = 0x8DBD
+const R3D_BC4: int = 0x8DBB
+function gvk_is_compressed(ifmt: int) -> bool { return ifmt == R3D_BC7 or ifmt == R3D_BC7_SRGB or ifmt == R3D_BC5 or ifmt == R3D_BC4 }
 function gvk_format(ifmt: int) -> int {
+  if ifmt == R3D_BC7 { return VK_FORMAT_BC7_UNORM_BLOCK }
+  if ifmt == R3D_BC7_SRGB { return VK_FORMAT_BC7_SRGB_BLOCK }
+  if ifmt == R3D_BC5 { return VK_FORMAT_BC5_UNORM_BLOCK }
+  if ifmt == R3D_BC4 { return VK_FORMAT_BC4_UNORM_BLOCK }
   if ifmt == GL_RGBA8 or ifmt == GL_RGB8 { return VK_FORMAT_R8G8B8A8_UNORM }
   if ifmt == GL_RGB10_A2 { return VK_FORMAT_A2B10G10R10_UNORM_PACK32 }   # the HDR10 screen and LDR image
   if ifmt == GL_SRGB8_ALPHA8 or ifmt == GL_SRGB8 { return VK_FORMAT_R8G8B8A8_SRGB }
@@ -158,7 +169,8 @@ function gvk_tex_storage(render3d_st: mut Render3dState, tex: int, array: bool,
   if samples < 1 { samples = 1 }
   if samples > 1 { levels = 1 }
   var usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT
-  if depth { usage = usage | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT } else { usage = usage | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT }
+  # a block-compressed image is only ever sampled and copied into: it cannot be drawn to
+  if depth { usage = usage | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT } else if not gvk_is_compressed(ifmt) { usage = usage | VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT }
   # DLSS reads and writes the frame from compute: a float colour target is storage too while
   # Streamline is running (8-bit sRGB formats cannot be, so only the float ones)
   if render3d_st.gsl_on and samples == 1 and (vkfmt == VK_FORMAT_R16G16B16A16_SFLOAT or vkfmt == VK_FORMAT_R32_SFLOAT) { usage = usage | VK_IMAGE_USAGE_STORAGE_BIT }
diff --git a/packages/ludic.render3d/r3d.ludic b/packages/ludic.render3d/r3d.ludic
index ae94eacb..695579ff 100644
--- a/packages/ludic.render3d/r3d.ludic
+++ b/packages/ludic.render3d/r3d.ludic
@@ -11,6 +11,8 @@ import "gpu.ludic"
 import "gpu_manifest.ludic"
 import "gpu_vk.ludic"
 import "gpu_vk_res.ludic"
+import "gpu_vk_bc.ludic"
+import "texture_dds.ludic"
 import "prof.ludic"
 import "drawstats.ludic"
 import "programs.ludic"
diff --git a/packages/ludic.render3d/texture.ludic b/packages/ludic.render3d/texture.ludic
index 8e6211ec..34275088 100644
--- a/packages/ludic.render3d/texture.ludic
+++ b/packages/ludic.render3d/texture.ludic
@@ -263,6 +263,18 @@ function tex_upload(render3d_st: mut Render3dState, px: pointer, srgb: bool, mip
 function tex_load(render3d_st: mut Render3dState, path: pointer, srgb: bool) -> int { return tex_load_ex(render3d_st, path, srgb, 0) }
 # ... with `dilate` passes of edge padding for a cut-out atlas (0 = none)
 function tex_load_ex(render3d_st: mut Render3dState, path: pointer, srgb: bool, dilate: int) -> int {
+  # a .dds encoded beside the .png at build time goes to the GPU compressed, its mips included
+  # (not for a cut-out atlas: its edge padding is made here, from the .png)
+  if render3d_st.gvk_has_bc and dilate == 0 {
+    let dds = dds_path_of(path)
+    if len(dds) > 0 and Fs.exists(dds) {
+      let t = tex_load_dds(render3d_st, dds, srgb)
+      if t != 0 {
+        tex_note_size(render3d_st, t)
+        return t
+      }
+    }
+  }
   let px = png_decode(render3d_st, path)
   if px == null { return 0 }
   if dilate > 0 { tex_dilate(render3d_st, px, 60, dilate) }
diff --git a/packages/ludic.render3d/texture_dds.ludic b/packages/ludic.render3d/texture_dds.ludic
new file mode 100644
index 00000000..dfdd6376
--- /dev/null
+++ b/packages/ludic.render3d/texture_dds.ludic
@@ -0,0 +1,71 @@
+# texture_dds.ludic - a .dds beside a .png: BC7 (colour, and anything packed into three channels),
+# BC5 (a normal's x and y) or BC4 (one channel), every mip level encoded at build time. Only the
+# DX10 form with its whole chain is taken; anything else is 0 and the caller loads the .png.
+
+# DXGI's numbers for the formats read here
+const DDS_BC4: int = 80
+const DDS_BC5: int = 83
+const DDS_BC7: int = 98
+const DDS_BC7_SRGB: int = 99
+
+function dds_u32(b: []byte, o: int) -> int { return b[o] | (b[o + 1] << 8) | (b[o + 2] << 16) | (b[o + 3] << 24) }
+
+# the .dds a .png may have beside it, or "" when the path is not a .png
+function dds_path_of(path: string) -> string {
+  let n = len(path)
+  if n < 5 or path[n - 4 .. n] != ".png" { return "" }
+  return path[0 .. n - 4] + ".dds"
+}
+
+# Load the .dds at `path` as a texture (srgb: colour, sampled as sRGB), or 0
+function tex_load_dds(render3d_st: mut Render3dState, path: string, srgb: bool) -> int {
+  let b = Fs.read_bytes(path)
+  if b == null { return 0 }
+  let id = tex_load_dds_bytes(render3d_st, b, srgb)
+  free(b)
+  if id == 0 { print(`r3d: {path} is not a DX10 .dds with its whole mip chain; the .png is used`) }
+  return id
+}
+
+function tex_load_dds_bytes(render3d_st: mut Render3dState, b: []byte, srgb: bool) -> int {
+  if len(b) < 148 or b[0] != 68 or b[1] != 68 or b[2] != 83 or b[3] != 32 { return 0 }
+  if b[84] != 68 or b[85] != 88 or b[86] != 49 or b[87] != 48 { return 0 }      # "DX10"
+  let h = dds_u32(b, 12)
+  let w = dds_u32(b, 16)
+  var levels = dds_u32(b, 28)
+  if levels < 1 { levels = 1 }
+  let dxgi = dds_u32(b, 128)
+  var ifmt = 0
+  var block = 16
+  if dxgi == DDS_BC7 or dxgi == DDS_BC7_SRGB { ifmt = R3D_BC7;  if srgb { ifmt = R3D_BC7_SRGB } }
+  if dxgi == DDS_BC5 { ifmt = R3D_BC5 }
+  if dxgi == DDS_BC4 { ifmt = R3D_BC4;  block = 8 }
+  if ifmt == 0 or w < 1 or h < 1 or levels != gvk_mip_levels(w, h) { return 0 }
+  let offs = words(levels + 1)
+  var at = 0
+  for l in 0 .. levels {
+    offs[l] = at
+    at += max((max(w >> l, 1) + 3) / 4, 1) * max((max(h >> l, 1) + 3) / 4, 1) * block
+  }
+  offs[levels] = at
+  if 148 + at > len(b) {
+    free(offs)
+    return 0
+  }
+  let id = gpu_tex_new(render3d_st)
+  gpu_tex_bind(render3d_st, GPU_TEX2D, id)
+  let ok = gpu_tex_compressed(render3d_st, ifmt, w, h, levels, mem_off(data_of(b), 148), offs)
+  free(offs)
+  if not ok {
+    gpu_tex_free(render3d_st, id)
+    return 0
+  }
+  gpu_tex_param(render3d_st, GPU_TEX2D, GL_TEXTURE_WRAP_S, GL_REPEAT)
+  gpu_tex_param(render3d_st, GPU_TEX2D, GL_TEXTURE_WRAP_T, GL_REPEAT)
+  gpu_tex_param(render3d_st, GPU_TEX2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR)
+  gpu_tex_param(render3d_st, GPU_TEX2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_LINEAR)
+  gpu_tex_paramf(render3d_st, GPU_TEX2D, GL_TEXTURE_MAX_ANISOTROPY_EXT, render3d_st.tex_anisotropy)
+  render3d_st.tex_w = w
+  render3d_st.tex_h = h
+  return id
+}
diff --git a/tools/ludic-cli/test.ludic b/tools/ludic-cli/test.ludic
index 75ba6609..f2a66f60 100644
--- a/tools/ludic-cli/test.ludic
+++ b/tools/ludic-cli/test.ludic
@@ -1053,6 +1053,7 @@ function cmd_dev_test() -> int {
   controller_case("state/component", "", " #10 big 10 1", "component.ludic (0.S: a component's header names its states; its getters, functions and events take them and the template never sees them)")
   headless_case("rendering/ui_render3d", "ok", "ui_render3d.ludic (ludic.ui's render3d backend builds against the renderer)")
   headless_line_case("rendering/release", "RELEASE OK", "release.ludic (model_release: meshes go, a shared texture stays until its last user goes)")
+  headless_line_case("rendering/bc", "BC OK", "bc.ludic (a .dds beside a .png arrives BC7-compressed with every mip level)")
   migrate_component_case()
   migrate_foreign_case()
   reject_case("rejected/runtime_type_clash", "PadButton is the runtime's enum", "a program's type named like one of the runtime's is refused")