AI 應用與模型部署
Firefox Smart Window Integrates Mistral Small 4, Letting the Browser Process Tab Context with Switchable Models
Mozilla has added Mistral Small 4 to the Smart Window beta in the United States and Canada, expanded the French-language version to France, and retained other model options. The 119B-parameter MoE model offers a long context window and reasoning modes, but zero data retention does not mean local inference, and the boundaries governing the transmission of tab content still require real-world testing.
Mozilla and Mistral announced on September 16 that the Firefox Smart Window beta will adopt Mistral Small 4. Users in the United States and Canada can select the new model, while France becomes the first new market with official French-language support. Smart Window is not a conventional sidebar chatbot. It uses tab context shared by the user to organize complex search journeys, retrieve previously viewed information, and provide citations. Mozilla emphasized that other models will remain available in different markets, indicating that it is attempting at the product layer to decouple browser context and retrieval services from model providers.
The underlying model is not a small on-device model. According to its public model card, Mistral Small 4 has 119B parameters and uses a Mixture-of-Experts (MoE) architecture with 128 experts, activating four experts per token and approximately 6.5B parameters in practice. It supports a 256K context window, text and image inputs, function calling, and per-request switching through `reasoning_effort`. Its weights are released under Apache 2.0, and the model can also be self-hosted with vLLM. However, Mozilla has not disclosed which quantization, inference region, context-truncation strategy, or reasoning mode Smart Window uses. These details directly affect time to first token, the amount of tab data transmitted externally, and multilingual quality.
The privacy commitments require precise interpretation: conversations are not stored on Mozilla’s servers by default, and Mistral promises zero data retention. This does not mean the model runs on the device, nor does it demonstrate that transmitted content is end-to-end encrypted or never processed by third parties. Engineering teams should next examine how the browser identifies which tabs are being sent, whether context can be revoked on a per-request basis, whether citations are traceable, and whether tool and conversation state remain consistent after switching models. Mozilla has also not published a Smart Window-specific evaluation, stating only that multilingual performance factored into its model selection. The claimed French-language tuning therefore cannot yet be treated as an independently verified quality improvement.