diff --git a/tests/models/multimodal/processing/test_glm5next.py b/tests/models/multimodal/processing/test_glm5next.py
index 719154a2b4..111fb982cf 100644
--- a/tests/models/multimodal/processing/test_glm5next.py
+++ b/tests/models/multimodal/processing/test_glm5next.py
@@ -188,3 +188,54 @@ def test_mm_device_do_normalize():
     torch.testing.assert_close(
         output, normalized_values.to(device), rtol=1e-5, atol=1e-6
     )
+
+
+def _image_info(**kwargs):
+    ctx = build_model_context(
+        "zai-org/GLM-5.3-Flash",
+        limit_mm_per_prompt={"image": 1},
+        **kwargs,
+    )
+    return MULTIMODAL_REGISTRY.create_processor(
+        ctx.model_config,
+        tokenizer=ctx.tokenizer,
+    ).info
+
+
+def test_image_encoder_cache_covers_full_token_budget():
+    """The most-features probe must reach the processor's token ceiling.
+
+    The inherited square probe refits to 89x89 = 7921 tokens under the
+    max_image_tokens=8000 budget, so the encoder cache came up short and
+    ordinary non-square images in 7922-8000 tokens were refused with
+    HTTP 400 (#59539).
+    """
+    info = _image_info()
+    assert info.get_max_image_tokens() == 8000
+
+    # The shapes from the issue, with the token counts the processor
+    # actually produces; every one must fit the cache.
+    for width, height, expected_tokens in [
+        (4032, 3024, 7931),  # phone photo, 4:3
+        (3840, 2160, 7973),  # 4K frame, 16:9
+        (3508, 2480, 7950),  # A4 at 300 dpi
+        (2600, 2400, 7998),  # 13:12
+        (3000, 3000, 7921),  # square worst case before the fix
+    ]:
+        num_tokens = info.get_num_image_tokens(image_width=width, image_height=height)
+        assert num_tokens == expected_tokens
+        assert num_tokens <= info.get_max_image_tokens()
+
+    # The profiling dummy covers the real worst case, not the square one.
+    size = info.get_image_size_with_most_features()
+    assert size.width * size.height == 2240 * 2800
+
+
+def test_image_encoder_cache_follows_max_pixels_override():
+    info = _image_info(mm_processor_kwargs={"max_pixels": 1568 * 100})
+    assert info.get_max_image_tokens() == 100
+    size = info.get_image_size_with_most_features()
+    assert (
+        info.get_num_image_tokens(image_width=size.width, image_height=size.height)
+        == 100
+    )
diff --git a/vllm/models/glm5next/common/multimodal.py b/vllm/models/glm5next/common/multimodal.py
index 75921b697f..7384c8c316 100644
--- a/vllm/models/glm5next/common/multimodal.py
+++ b/vllm/models/glm5next/common/multimodal.py
@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """GLM-5.3-Flash vision tower and multimodal processor."""
 
+import math
 from collections.abc import Mapping
 from functools import partial
 from typing import Any
@@ -672,6 +673,29 @@ class Glm5NextProcessingInfo(Glm4vProcessingInfo):
             return int(override)
         return self._processor_pixel_budget(self.get_hf_processor().video_processor)[1]
 
+    def get_image_size_with_most_features(self) -> ImageSize:
+        # The inherited square probe strands budget whenever the token
+        # ceiling is not a perfect square: with max_image_tokens=8000 the
+        # square refits to 2492x2492 (89x89 = 7921 tokens) while a
+        # 2240x2800 canvas reaches 80x100 = 8000, so the encoder cache came
+        # up short of the processor's own maximum and refused valid images
+        # (#59539). Factor the ceiling exactly instead of probing a square.
+        vision_config = self.get_hf_config().vision_config
+        factor = (
+            vision_config.patch_size
+            * vision_config.spatial_merge_size
+            * self.get_hf_processor().image_processor.patch_expand_factor
+        )
+        pixels_per_token = vision_config.temporal_patch_size * factor * factor
+        max_tokens = max(1, self._get_image_max_pixels() // pixels_per_token)
+        short_side = math.isqrt(max_tokens)
+        while max_tokens % short_side:
+            short_side -= 1
+        return ImageSize(
+            width=(max_tokens // short_side) * factor,
+            height=short_side * factor,
+        )
+
     def _get_video_second_idx_glm46v(
         self, metadata: dict[str, Any], total_frames: int
     ) -> list[int]: