diff --git a/packages/core/src/models/modelsConfig.test.ts b/packages/core/src/models/modelsConfig.test.ts
index 1159f639238..9d5434b52fa 100644
--- a/packages/core/src/models/modelsConfig.test.ts
+++ b/packages/core/src/models/modelsConfig.test.ts
@@ -4,7 +4,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-import { describe, it, expect } from 'vitest';
+import { describe, it, expect, vi } from 'vitest';
 import { ModelsConfig } from './modelsConfig.js';
 import { AuthType } from '../core/contentGenerator.js';
 import type { ContentGeneratorConfig } from '../core/contentGenerator.js';
@@ -1331,6 +1331,123 @@ describe('ModelsConfig', () => {
     expect(modelsConfig.getGenerationConfig().model).toBe('custom-model');
   });
 
+  it('recomputes raw model modalities instead of carrying provider multimodal defaults', async () => {
+    const modelProvidersConfig: ModelProvidersConfig = {
+      openai: [
+        {
+          id: 'qwen3.6-plus',
+          name: 'Qwen 3.6 Plus',
+          baseUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
+          envKey: 'DASHSCOPE_API_KEY',
+          generationConfig: {
+            contextWindowSize: 12345,
+            modalities: { image: true, video: true },
+          },
+        },
+      ],
+    };
+
+    const modelsConfig = new ModelsConfig({
+      initialAuthType: AuthType.USE_OPENAI,
+      modelProvidersConfig,
+    });
+
+    await modelsConfig.switchModel(AuthType.USE_OPENAI, 'qwen3.6-plus');
+    expect(modelsConfig.getGenerationConfig().modalities).toEqual({
+      image: true,
+      video: true,
+    });
+
+    await modelsConfig.setModel('qwen3.7-max');
+
+    expect(modelsConfig.getModel()).toBe('qwen3.7-max');
+    expect(modelsConfig.getGenerationConfig().modalities).toEqual({});
+    expect(modelsConfig.getGenerationConfigSources()['modalities']).toEqual({
+      kind: 'computed',
+      detail: 'auto-detected from model',
+    });
+    expect(modelsConfig.getGenerationConfig().contextWindowSize).not.toBe(
+      12345,
+    );
+    expect(
+      modelsConfig.getGenerationConfigSources()['contextWindowSize'],
+    ).toEqual({
+      kind: 'computed',
+      detail: 'auto-detected from model',
+    });
+  });
+
+  it('notifies the owner to refresh after a raw model switch', async () => {
+    const onModelChange = vi.fn();
+    const modelsConfig = new ModelsConfig({
+      initialAuthType: AuthType.USE_OPENAI,
+      generationConfig: {
+        model: 'qwen3.6-plus',
+        modalities: { image: true, video: true },
+      },
+      onModelChange,
+    });
+
+    await modelsConfig.setModel('qwen3.7-max');
+
+    expect(onModelChange).toHaveBeenCalledWith(AuthType.USE_OPENAI, true);
+  });
+
+  it('preserves explicitly configured modalities during raw model switches', async () => {
+    const modelsConfig = new ModelsConfig({
+      initialAuthType: AuthType.USE_OPENAI,
+      generationConfig: {
+        model: 'custom-vision-model',
+        modalities: { image: true },
+      },
+      generationConfigSources: {
+        modalities: {
+          kind: 'settings',
+          settingsPath: 'model.generationConfig.modalities',
+        },
+      },
+    });
+
+    await modelsConfig.setModel('custom-vision-model-v2');
+
+    expect(modelsConfig.getGenerationConfig().modalities).toEqual({
+      image: true,
+    });
+    expect(modelsConfig.getGenerationConfigSources()['modalities']).toEqual({
+      kind: 'settings',
+      settingsPath: 'model.generationConfig.modalities',
+    });
+  });
+
+  it('rolls back raw model state when owner refresh fails', async () => {
+    const modelsConfig = new ModelsConfig({
+      initialAuthType: AuthType.USE_OPENAI,
+      generationConfig: {
+        model: 'qwen3.6-plus',
+        modalities: { image: true, video: true },
+      },
+      generationConfigSources: {
+        modalities: {
+          kind: 'computed',
+          detail: 'auto-detected from model',
+        },
+      },
+      onModelChange: async () => {
+        throw new Error('refresh failed');
+      },
+    });
+
+    await expect(modelsConfig.setModel('qwen3.7-max')).rejects.toThrow(
+      'refresh failed',
+    );
+
+    expect(modelsConfig.getModel()).toBe('qwen3.6-plus');
+    expect(modelsConfig.getGenerationConfig().modalities).toEqual({
+      image: true,
+      video: true,
+    });
+  });
+
   it('should maintain consistency between currentModelId and _generationConfig.model during updateCredentials', () => {
     const modelsConfig = new ModelsConfig({
       initialAuthType: AuthType.USE_OPENAI,
diff --git a/packages/core/src/models/modelsConfig.ts b/packages/core/src/models/modelsConfig.ts
index 0c90931905b..30a8f236ded 100644
--- a/packages/core/src/models/modelsConfig.ts
+++ b/packages/core/src/models/modelsConfig.ts
@@ -350,12 +350,61 @@ export class ModelsConfig {
     }
 
     // Raw model override: update generation config in-place
-    this.strictModelProviderSelection = false;
-    this._generationConfig.model = newModel;
-    this.generationConfigSources['model'] = {
-      kind: 'programmatic',
-      detail: metadata?.reason || 'setModel',
-    };
+    const rollbackSnapshot = this.createStateSnapshotForRollback();
+    try {
+      this.strictModelProviderSelection = false;
+      this._generationConfig.model = newModel;
+      this.generationConfigSources['model'] = {
+        kind: 'programmatic',
+        detail: metadata?.reason || 'setModel',
+      };
+      this.applyRawModelDerivedDefaults(newModel);
+
+      if (this.onModelChange && this.currentAuthType) {
+        await this.onModelChange(this.currentAuthType, true);
+      }
+    } catch (error) {
+      this.rollbackToStateSnapshot(rollbackSnapshot);
+      throw error;
+    }
+  }
+
+  /**
+   * Raw model switches keep the current credentials, but model-derived
+   * generation defaults must follow the new model. Otherwise a switch from a
+   * multimodal registry model to a text-only raw model can keep stale image
+   * support and send unsupported inline media.
+   */
+  private applyRawModelDerivedDefaults(modelId: string): void {
+    if (this.shouldUpdateModelDerivedDefault('modalities')) {
+      this._generationConfig.modalities = defaultModalities(modelId);
+      this.generationConfigSources['modalities'] = {
+        kind: 'computed',
+        detail: 'auto-detected from model',
+      };
+    }
+
+    if (this.shouldUpdateModelDerivedDefault('contextWindowSize')) {
+      this._generationConfig.contextWindowSize = tokenLimit(modelId, 'input');
+      this.generationConfigSources['contextWindowSize'] = {
+        kind: 'computed',
+        detail: 'auto-detected from model',
+      };
+    }
+  }
+
+  private shouldUpdateModelDerivedDefault(
+    field: 'modalities' | 'contextWindowSize',
+  ): boolean {
+    const source = this.generationConfigSources[field];
+    return (
+      source === undefined ||
+      source.kind === 'computed' ||
+      source.kind === 'default' ||
+      source.kind === 'modelProviders' ||
+      source.kind === 'programmatic' ||
+      source.kind === 'unknown'
+    );
   }
 
   /**