The Vercel AI SDK has released an update for its MoonshotAI provider, @ai-sdk/moonshotai@2.0.45, which now includes support for video input. This version also signifies a change in the provider's architecture, as it now owns its chat implementation, converter, language model, and helpers, rather than building on @ai-sdk/openai-compatible.
Key Points
- The
@ai-sdk/moonshotaiprovider now supports video input, converting video file parts (e.g.,mediaType: 'video/mp4') to Moonshot'svideo_urlcontent parts. - This video input capability is available for video-capable Moonshot models, including kimi-k3, kimi-k2.7-code, kimi-k2.6, and kimi-k2.5.
- The provider's chat implementation, converter, language model, and helpers are now owned directly by the
@ai-sdk/moonshotaipackage. - Audio and PDF file parts will now trigger client-side errors, as the Moonshot API rejects these part types.
- The
reasoningHistory: 'preserved'option now maps to Moonshot'sthinking.keep: 'all'request field, with a warning for models that do not supportthinking.keep. - New provider options include
promptCacheKeyandsafetyIdentifier, andreasoningEffortnow accepts'low','high', or'max'values. - The
@ai-sdk/xaiprovider also added support for the Grok 4.6 model IDs and itsxhighreasoning effort in version@ai-sdk/xai@4.0.37, and the Grok 4.6 models in version@ai-sdk/xai@2.0.86.
Context
According to the Vercel AI SDK changelog, the @ai-sdk/moonshotai update streamlines the integration with Moonshot's capabilities by directly managing its core components. This change allows for specific feature implementations like video input, which is then translated into Moonshot's native video_url format for compatible models. The changelog also indicates that the @ai-sdk/xai provider has introduced support for the Grok 4.6 model and its xhigh reasoning effort, as well as a priority service tier for chat and responses.
Why It Matters
This update provides developers using the Vercel AI SDK with direct access to Moonshot's multimodal capabilities, specifically video input, and offers more granular control over reasoning efforts. The shift to an owned chat implementation suggests a more tailored and potentially optimized integration for Moonshot models, while the Grok updates expand the range of supported models and features for XAI users.
What To Do
- Developers should review the updated
@ai-sdk/moonshotaidocumentation for details on implementing video input. - Note the new
reasoningEffortvalues andpromptCacheKeyandsafetyIdentifieroptions for Moonshot models. - Test existing applications for compatibility, especially regarding audio and PDF file parts, which now throw client-side errors.
- For XAI users, explore the new priority service tier and the Grok 4.6 model capabilities.
Keep Exploring
/atlas/**grok**-family /atlas/image-models /techniques/multimodal-grounding /studio?pack=foundation
