diff --git a/benchmark/timeseries/gift_eval/README.md b/benchmark/timeseries/gift_eval/README.md
index d49f394c0..b0117fad6 100644
--- a/benchmark/timeseries/gift_eval/README.md
+++ b/benchmark/timeseries/gift_eval/README.md
@@ -13,14 +13,18 @@ pip install --no-deps "git+https://github.com/SalesforceAIResearch/gift-eval.git
 
 ## Run
 
+- **`TimesFM3`**:
+
+  ```bash
+  python -m benchmark.timeseries.gift_eval.main --model timesfm3
+  ```
+
+  `TimesFM3` follows official evaluation with symmetric averaging over two estimators, contexts truncated to 15,360 time steps and up to 32 variates per forward pass.
+
 > [!NOTE]
 > Weights of `TimesFM3` are distributed under the [TimesFM Non-Commercial License v1.0](https://huggingface.co/google/timesfm-3.0-pytorch/blob/main/LICENSE).
 > Review the license before running the `TimesFM3` benchmark, which will download its weights noninteractively.
 
-```bash
-python -m benchmark.timeseries.gift_eval.main --model timesfm3
-```
-
 Results will be written to `benchmark/timeseries/gift_eval_out/<model>/`, and already completed configurations are skipped.
 
 ## Evaluate
diff --git a/benchmark/timeseries/gift_eval/predictor.py b/benchmark/timeseries/gift_eval/predictor.py
index 4266a5517..52d801bc9 100644
--- a/benchmark/timeseries/gift_eval/predictor.py
+++ b/benchmark/timeseries/gift_eval/predictor.py
@@ -36,6 +36,8 @@ def predict(
     ) -> Iterator[QuantileForecast]:
 
         for batch in batcher(dataset, batch_size=self.batch_size):
+            cpu_generator = torch.Generator().manual_seed(42)
+
             ys = [torch.from_numpy(entry["target"]) for entry in batch]
             ys = [torch.atleast_2d(y) for y in ys]
             xs = []
@@ -67,6 +69,13 @@ def predict(
             for i, x_i in enumerate(xs):
                 x[i, :, -x_i.size(-1) :] = x_i
 
+            if (
+                self.config.max_variates is not None
+                and x.size(1) > self.config.max_variates - 1
+            ):  # Keep room for at least one target:
+                perm = torch.randperm(x.size(1), generator=cpu_generator)
+                x = x[:, perm[: self.config.max_variates - 1]]
+
             y = y.transpose(-1, -2).to(
                 self.device,
                 memory_format=torch.contiguous_format,
@@ -76,17 +85,30 @@ def predict(
                 memory_format=torch.contiguous_format,
             )
 
-            out = self.model(
-                x_context=x,
-                y_context=y,
-                x_query=torch.empty(
-                    (x.size(0), self.prediction_length, 0),
-                    device=self.device,
-                ),
-                num_estimators=self.config.num_estimators,
-            )
-            pred = out.numerical
-            pred = pred.view(y.size(0), self.prediction_length, y.size(2), -1)
+            size = y.size(-1)
+            if self.config.max_variates is not None:
+                size = min(size, self.config.max_variates - x.size(-1))
+
+            preds = []
+            for start in range(0, y.size(-1), size):
+                y_i = y[..., start : start + size]
+                if y_i.size(-1) < size:
+                    y_i = torch.cat([y_i, y[..., : size - y_i.size(-1)]], -1)
+
+                out = self.model(
+                    x_context=x,
+                    y_context=y_i,
+                    x_query=torch.empty(
+                        (x.size(0), self.prediction_length, 0),
+                        device=self.device,
+                    ),
+                    num_estimators=self.config.num_estimators,
+                )
+                pred = out.numerical
+                pred = pred.view(y.size(0), self.prediction_length, size, -1)
+                preds.append(pred)
+
+            pred = torch.cat(preds, dim=-2)[..., : y.size(-1), :]
             pred = pred.permute(0, 3, 1, 2).cpu().numpy()  # [B, Q, R, C]
             levels = out.numerical_columns[: pred.shape[1]]
             levels = [str(float(f"0.{q.split('q')[-1]}")) for q in levels]
diff --git a/benchmark/timeseries/model.py b/benchmark/timeseries/model.py
index feb9b85db..2689b66aa 100644
--- a/benchmark/timeseries/model.py
+++ b/benchmark/timeseries/model.py
@@ -15,6 +15,7 @@ class ModelConfig:
     factory: Callable[[torch.device], sdm.models.ICLModel]
     num_estimators: int = 1
     max_context_size: int | None = None
+    max_variates: int | None = None
 
 
 MODEL_CONFIGS = {
@@ -26,5 +27,6 @@ class ModelConfig:
         ),
         num_estimators=2,
         max_context_size=15_360,
+        max_variates=32,
     ),
 }