diff --git a/Tests/MLXLMTests/VLMGemma4Tests.swift b/Tests/MLXLMTests/VLMGemma4Tests.swift
new file mode 100644
index 000000000..1a0962040
--- /dev/null
+++ b/Tests/MLXLMTests/VLMGemma4Tests.swift
@@ -0,0 +1,133 @@
+import Foundation
+import MLX
+import MLXLMCommon
+import MLXNN
+import MLXVLM
+import Testing
+
+/// The first unit tests for anything under MLXVLM.
+///
+/// Measured coverage put the whole VLM half of the model layer at 0.3% over 13,304
+/// lines, with `MLXVLM/Models/Gemma4.swift` — 1,771 lines — the largest file no test
+/// executed. That file is where SharpAI/mlx-swift-lm#45's changes landed, which is
+/// what issue #128 flagged and what this starts to close.
+///
+/// Vision models are exercised end to end by the `vision` and `omni` CI jobs, so the
+/// gap is not that they never run. It is that a shape or dtype fault inside one
+/// surfaces as a wrong answer about an image rather than a failing assertion, which
+/// is a slow and ambiguous way to find out.
+///
+/// These construct from tiny configurations and run a forward pass. They establish
+/// shape and finiteness, not numerical correctness — the weights are random.
+extension MLXTestingSuite {
+    @Suite
+    struct VLMGemma4Tests {
+
+        /// Mirrors the shape of a real gemma-4 VLM config: a text half and a vision
+        /// half, each with its own model_type, under one wrapper.
+        private func tinyConfigData() -> Data {
+            let json = """
+            {
+                "model_type": "gemma4",
+                "image_token_id": 3,
+                "audio_token_id": 4,
+                "boi_token_id": 5,
+                "eoi_token_id": 6,
+                "vision_soft_tokens_per_image": 4,
+                "text_config": {
+                    "model_type": "gemma4_text",
+                    "hidden_size": 64,
+                    "num_hidden_layers": 2,
+                    "intermediate_size": 128,
+                    "num_attention_heads": 4,
+                    "num_key_value_heads": 2,
+                    "head_dim": 16,
+                    "global_head_dim": 16,
+                    "rms_norm_eps": 1e-6,
+                    "vocab_size": 128,
+                    "vocab_size_per_layer_input": 16,
+                    "num_kv_shared_layers": 0,
+                    "hidden_size_per_layer_input": 32,
+                    "sliding_window": 128,
+                    "sliding_window_pattern": 1,
+                    "max_position_embeddings": 512,
+                    "rope_traditional": false,
+                    "rope_theta": 10000.0,
+                    "use_double_wide_mlp": false,
+                    "tie_word_embeddings": true,
+                    "final_logit_softcapping": 30.0,
+                    "enable_moe_block": false,
+                    "attention_k_eq_v": false
+                },
+                "vision_config": {
+                    "model_type": "gemma4_vision",
+                    "num_hidden_layers": 2,
+                    "hidden_size": 32,
+                    "intermediate_size": 64,
+                    "num_attention_heads": 4,
+                    "num_key_value_heads": 4,
+                    "head_dim": 8,
+                    "patch_size": 16,
+                    "rms_norm_eps": 1e-6,
+                    "default_output_length": 4,
+                    "position_embedding_size": 16,
+                    "pooling_kernel_size": 2,
+                    "use_clipped_linears": false,
+                    "standardize": false,
+                    "rope_parameters": {"rope_theta": 10000.0}
+                }
+            }
+            """
+            return json.data(using: .utf8)!
+        }
+
+        @Test("VLM Gemma4 configuration decodes both halves")
+        func testConfigurationDecoding() throws {
+            let config = try JSONDecoder().decode(
+                Gemma4Configuration.self, from: tinyConfigData())
+
+            // The wrapper derives these from the text half when absent, which is the
+            // behaviour a checkpoint without top-level copies depends on.
+            #expect(config.vocabularySize == 128)
+            #expect(config.hiddenSize == 64)
+            #expect(config.imageTokenId == 3)
+        }
+
+        @Test("VLM Gemma4 instantiates from a tiny configuration")
+        func testInstantiation() throws {
+            let config = try JSONDecoder().decode(
+                Gemma4Configuration.self, from: tinyConfigData())
+            let model = Gemma4(config)
+            #expect(model.vocabularySize == 128)
+        }
+
+        @Test("VLM Gemma4 text-only forward pass has the right shape")
+        func testTextOnlyForwardPass() throws {
+            let config = try JSONDecoder().decode(
+                Gemma4Configuration.self, from: tinyConfigData())
+            let model = Gemma4(config)
+
+            // No image: a VLM still has to answer a text-only prompt, and that path is
+            // what every chat request without an attachment takes.
+            let input = LMInput(tokens: MLXArray(0 ..< 6))
+            let result = model(input.text.tokens[.newAxis], cache: nil)
+
+            #expect(result.shape == [1, 6, 128])
+            let sum = result.sum().item(Float.self)
+            #expect(!sum.isNaN)
+            #expect(!sum.isInfinite)
+        }
+
+        @Test("VLM Gemma4 forward pass is deterministic")
+        func testDeterminism() throws {
+            let config = try JSONDecoder().decode(
+                Gemma4Configuration.self, from: tinyConfigData())
+            let model = Gemma4(config)
+
+            let input = MLXArray(0 ..< 5).reshaped(1, 5)
+            let a = model(input, cache: nil)
+            let b = model(input, cache: nil)
+            #expect(allClose(a, b).item(Bool.self))
+        }
+    }
+}