diff --git a/app/ui/features/generate/warn.js b/app/ui/features/generate/warn.js
index c525cf5e..597c14a0 100644
--- a/app/ui/features/generate/warn.js
+++ b/app/ui/features/generate/warn.js
@@ -1,4 +1,5 @@
-// The chips beside Generate: the pipeline has nowhere to put the scene text; a learning feature is on with no Taste key.
+// The chips beside Generate: the pipeline has nowhere to put the scene text; a learning feature is on with no Taste key;
+// this torch build runs int8 models slowly.
 import { composer as c } from "../../composer/composer.js";
 
 export default {
@@ -20,7 +21,13 @@ export default {
     const twice = c.button.sm({ label: "⚠ Two sparse attentions", tone: "ghost", disabled: true,
       title: "SLA and ComfyUI's Model Sparse Attention node are both in this pipeline. Both sparsify attention; remove one." }).node;
     twice.hidden = true;
-    host.append(chip, plain, keyless, twice);
+    // A torch build that halves int8 speed on this GPU (cu12x on Blackwell). Asked once: the build cannot change while ComfyUI runs.
+    const slow = c.button.sm({ label: "⚠ Slow torch for int8", tone: "ghost", disabled: true, title: "" }).node;
+    slow.hidden = true;
+    host.append(chip, plain, keyless, twice, slow);
+    Promise.resolve().then(() => app.api?.torchBuild?.()).then((r) => {
+      if (r && r.slow_int8) { slow.title = r.slow_int8; slow.hidden = false; }
+    }).catch(() => {});      // unknown build: the loader still says it in the log when an int8 model loads
     const draw = (slots) => { chip.hidden = !(slots && !slots.some((s) => (s.roles || []).some((r) => r.at === "generation.prompt"))); };
     const drawTwice = (slots) => {
       const native = (slots || []).some((s) => /SparseAttention/.test(String(s.node || "")));
diff --git a/app/ui/features/generate/warn.test.js b/app/ui/features/generate/warn.test.js
index be70ee33..62fd4c14 100644
--- a/app/ui/features/generate/warn.test.js
+++ b/app/ui/features/generate/warn.test.js
@@ -35,3 +35,23 @@ test("the two-sparse-attentions chip shows only when SLA and a native sparse nod
   listener();
   assert.equal(twice().hidden, true, "SLA off: the native node alone is not a clash");
 });
+
+test("the slow-torch chip shows the server's sentence when this build runs int8 slowly, and stays hidden otherwise", async () => {
+  const pipeline = { slots: () => [], allOff: () => false, modulesById: () => ({}), activeModules: () => [], useful: () => false,
+    currentValues: () => ({}), subscribe: () => () => {} };
+  const slowChip = (host) => [...host.querySelectorAll("button")].find((b) => /Slow torch/.test(b.textContent));
+  const tick = () => new Promise((r) => setTimeout(r, 0));
+  let host = document.createElement("div");
+  warn.setup({ host, app: { pipeline, api: { torchBuild: async () => ({ slow_int8: "torch is built for CUDA 12.8: reinstall" }) } } });
+  await tick();
+  assert.equal(slowChip(host).hidden, false);
+  assert.match(slowChip(host).title, /CUDA 12\.8/);
+  host = document.createElement("div");
+  warn.setup({ host, app: { pipeline, api: { torchBuild: async () => ({ slow_int8: null }) } } });
+  await tick();
+  assert.equal(slowChip(host).hidden, true, "a good build says nothing");
+  host = document.createElement("div");
+  warn.setup({ host, app: { pipeline, api: { torchBuild: async () => { throw new Error("offline"); } } } });
+  await tick();
+  assert.equal(slowChip(host).hidden, true, "an unreadable build is not a claim");
+});
diff --git a/app/ui/shell/api.js b/app/ui/shell/api.js
index 0e64440b..4160892f 100644
--- a/app/ui/shell/api.js
+++ b/app/ui/shell/api.js
@@ -54,6 +54,7 @@ export const api = {
   shotMemory: () => call("GET", "/api/m/conditioning_shot_camera/memory"),
   forgetShotMemory: (kind, name) => call("POST", "/api/m/conditioning_shot_camera/forget", { kind, name }),
   tasteKeys: () => call("GET", "/api/m/taste/keys"),
+  torchBuild: () => call("GET", "/api/m/loader_diffusion_model/torch"),
   blockInfluence: (key) => call("GET", `/api/m/block_influence/status?key=${encodeURIComponent(key || "default")}`),
   blockInfluenceGroups: (key) => call("GET", `/api/m/block_influence/groups?key=${encodeURIComponent(key || "default")}`),
   clearBlockInfluence: (key) => call("POST", "/api/m/block_influence/clear", { key: key || "default" }),
diff --git a/modules/loaders/common.py b/modules/loaders/common.py
index 2c9ab3bd..6dfc6428 100644
--- a/modules/loaders/common.py
+++ b/modules/loaders/common.py
@@ -49,6 +49,42 @@ def unsupported_dtypes(model, weight_dtype, compute_dtype):
             if picked in ("fp16", "bf16", "fp32") and supported and dtype_of(picked) not in supported]
 
 
+_SHORT = {"bfloat16": "bf16", "float16": "fp16", "float32": "fp32"}
+
+
+def stored_weights(state_dict):
+    """What a file's weights are stored as, by element count: "bf16", or "int8+bf16" for a
+    mixed file. Read from the tensors, not the weight_dtype widget: a pre-quantized file
+    loaded with weight_dtype bf16 still runs int8 layers, and the log must say so."""
+    sizes = {}
+    for t in state_dict.values():
+        if torch.is_tensor(t) and t.dim() >= 2:
+            name = str(t.dtype).replace("torch.", "")
+            sizes[_SHORT.get(name, name)] = sizes.get(_SHORT.get(name, name), 0) + t.numel()
+    total = sum(sizes.values())
+    return "+".join(n for n in sorted(sizes, key=sizes.get, reverse=True) if sizes[n] > total * 0.1) or None
+
+
+def slow_int8_build():
+    """Why int8 models sample slowly on this torch build, or None.
+
+    Measured on an RTX PRO 6000 Blackwell, same 4-step H3 run: torch 2.10+cu128 sampled the
+    int8 model ~2x SLOWER than bf16; the same torch from the cu130 index made int8 ~25% faster.
+    """
+    cuda = getattr(torch.version, "cuda", None)
+    try:
+        if not cuda or not torch.cuda.is_available() or torch.cuda.get_device_capability()[0] < 10:
+            return None
+        if int(cuda.split(".")[0]) >= 13:
+            return None
+    except (RuntimeError, ValueError):
+        return None
+    # --force-reinstall is required: pip counts 2.10.0+cu128 as already satisfying torch==2.10.0.
+    return (f"torch is built for CUDA {cuda}: on this Blackwell GPU int8 models sample about 2x slower "
+            f"than on a CUDA 13 build. Stop ComfyUI and reinstall torch, torchvision and torchaudio "
+            f"(same versions) with --force-reinstall --index-url https://download.pytorch.org/whl/cu130")
+
+
 def weight_model_options(weight_dtype):
     """comfy `model_options` for a weight dtype. The mapping core's UNETLoader uses."""
     options = {}
diff --git a/modules/loaders/diffusion_model/__init__.py b/modules/loaders/diffusion_model/__init__.py
index 09f8cd34..0f1761d8 100644
--- a/modules/loaders/diffusion_model/__init__.py
+++ b/modules/loaders/diffusion_model/__init__.py
@@ -9,3 +9,14 @@ CATEGORY = "system"
 STATUS = "proven"
 
 NODES = [FunPackDiffusionModelLoader]
+
+
+def routes(table, base, web):
+    @table.get(base + "/torch")
+    async def _torch(_req):
+        """Why int8 would run slow on this torch build (None when it would not), for the chip beside Generate."""
+        from ..common import slow_int8_build
+        return web.json_response({"slow_int8": slow_int8_build()})
+
+
+PROVIDES = {"routes": routes}
diff --git a/modules/loaders/diffusion_model/nodes.py b/modules/loaders/diffusion_model/nodes.py
index 89506183..71144510 100644
--- a/modules/loaders/diffusion_model/nodes.py
+++ b/modules/loaders/diffusion_model/nodes.py
@@ -16,7 +16,7 @@ from .. import gguf_support, sla_attention
 from .. import int8_convrot as int8_convrot_module
 from ..common import (COMPUTE_DTYPES, WEIGHT_DTYPES, attention_choices, unsupported_dtypes,
                       attention_override, dtype_of, set_fp16_accumulation,
-                      weight_model_options)
+                      slow_int8_build, stored_weights, weight_model_options)
 
 
 def _model_file_choices():
@@ -230,6 +230,13 @@ class FunPackDiffusionModelLoader(io.ComfyNode):
                 notes.append(note)
                 if compute_dtype != "default":
                     notes.append("int8_convrot: compute dtype is forced, so the matmuls run in that dtype -- memory is saved, speed is not")
+        # Read before the load: comfy may consume the dict.
+        gguf = gguf_support.is_gguf(model_name) or misnamed
+        stored = None if gguf else stored_weights(state_dict)
+        slow = slow_int8_build() if stored and "int8" in stored else None
+        if slow:
+            log.alert("FunPack Diffusion Model Loader", slow)
+            notes.append(slow)
         model = comfy.sd.load_diffusion_model_state_dict(
             state_dict, model_options=model_options, metadata=metadata)
         if model is None:
@@ -290,8 +297,11 @@ class FunPackDiffusionModelLoader(io.ComfyNode):
         kind = type(getattr(model, "model", model)).__name__
         applied_attention = "sla_h3" if installed else (
             attention if override is not None else "default (as launched)")
+        # The file's own storage, not the widget: an int8 file loaded at "bf16" still runs int8.
+        weights = weight_dtype if not stored else (
+            stored if weight_dtype in ("default", stored) else f"{stored} in file ({weight_dtype} asked)")
         log.info("FunPack Diffusion Model Loader",
                  f"{model_name} loaded as {kind}, weights "
-                 f"{weight_dtype}, compute {compute_dtype}, attention "
+                 f"{weights}, compute {compute_dtype}, attention "
                  f"{applied_attention}")
         return io.NodeOutput(model, "\n".join(notes))
diff --git a/modules/loaders/tests/test_weight_report.py b/modules/loaders/tests/test_weight_report.py
new file mode 100644
index 00000000..86b4652c
--- /dev/null
+++ b/modules/loaders/tests/test_weight_report.py
@@ -0,0 +1,73 @@
+"""The loader says what the FILE holds, and when this torch build makes int8 slow."""
+
+import pytest
+import torch
+
+
+@pytest.fixture(autouse=True)
+def _needs_comfy(comfyui):
+    """The loader test imports a node module, which imports comfy."""
+
+
+def test_stored_weights_reads_the_tensors_not_the_widget():
+    from modules.loaders.common import stored_weights
+    mixed = {"a": torch.zeros(64, 64, dtype=torch.int8), "b": torch.zeros(32, 32, dtype=torch.bfloat16),
+             "scale": torch.zeros(64, dtype=torch.float32), "tiny": torch.zeros(2, 2, dtype=torch.float32)}
+    assert stored_weights(mixed) == "int8+bf16", "1-D scales and a <10% sliver are not the format"
+    assert stored_weights({"a": torch.zeros(8, 8, dtype=torch.bfloat16)}) == "bf16"
+    assert stored_weights({"n": 1}) is None
+
+
+@pytest.mark.parametrize("cuda,cap,gpu,slow", [
+    ("12.8", (12, 0), True, True),      # the rental that ran int8 2x slower
+    ("13.0", (12, 0), True, False),
+    ("12.8", (8, 9), True, False),      # not Blackwell: not measured, not claimed
+    ("12.8", (12, 0), False, False),
+    (None, (12, 0), True, False),       # CPU / MPS build
+])
+def test_slow_int8_build_names_only_cu12_on_blackwell(monkeypatch, cuda, cap, gpu, slow):
+    from modules.loaders import common
+    monkeypatch.setattr(torch.version, "cuda", cuda)
+    monkeypatch.setattr(torch.cuda, "is_available", lambda: gpu)
+    monkeypatch.setattr(torch.cuda, "get_device_capability", lambda *a: cap)
+    msg = common.slow_int8_build()
+    assert bool(msg) is slow
+    if slow:
+        assert "--force-reinstall" in msg and "cu130" in msg, "without the flag pip keeps the cu128 build"
+
+
+def _load(monkeypatch, sd, weight_dtype, slow=None):
+    import comfy.sd
+    import comfy.utils
+    import folder_paths
+    from modules.loaders import gguf_support
+    from modules.loaders.diffusion_model import nodes
+    monkeypatch.setattr(folder_paths, "get_full_path_or_raise", lambda kind, name: "/fake/path")
+    monkeypatch.setattr(gguf_support, "has_gguf_magic", lambda path: False)
+    monkeypatch.setattr(comfy.utils, "load_torch_file", lambda path, return_metadata: (sd, {}))
+    monkeypatch.setattr(comfy.sd, "load_diffusion_model_state_dict", lambda sd, model_options, metadata: object())
+    monkeypatch.setattr(nodes, "slow_int8_build", lambda: slow)
+    said = []
+    monkeypatch.setattr(nodes.log, "info", lambda src, msg: said.append(("info", msg)))
+    monkeypatch.setattr(nodes.log, "alert", lambda src, msg: said.append(("alert", msg)))
+    nodes.FunPackDiffusionModelLoader.execute(model_name="model.safetensors", weight_dtype=weight_dtype,
+                                              compute_dtype="bf16", attention="default")
+    return said
+
+
+def test_an_int8_file_loaded_at_bf16_is_logged_as_int8(monkeypatch):
+    said = _load(monkeypatch, {"w": torch.zeros(8, 8, dtype=torch.int8)}, "bf16")
+    line = [m for lvl, m in said if lvl == "info" and "loaded as" in m][0]
+    assert "weights int8 in file (bf16 asked)" in line
+
+
+def test_a_plain_file_reads_as_before(monkeypatch):
+    said = _load(monkeypatch, {"w": torch.zeros(8, 8, dtype=torch.bfloat16)}, "bf16", slow="slow build")
+    line = [m for lvl, m in said if lvl == "info" and "loaded as" in m][0]
+    assert "weights bf16," in line
+    assert not [m for lvl, m in said if lvl == "alert"], "the slow-build alert is about int8 files only"
+
+
+def test_an_int8_file_on_a_slow_build_raises_the_alert(monkeypatch):
+    said = _load(monkeypatch, {"w": torch.zeros(8, 8, dtype=torch.int8)}, "default", slow="slow build")
+    assert ("alert", "slow build") in said