Vercel AI SDK Releases @ai-sdk/xai 2.0.94 Patch Update

Vercel's AI SDK ecosystem has released a new version: @ai-sdk/xai@2.0.94. This is a patch-level release that addresses a specific issue in reasoning stream handling. While the version bump is minor, for developers who rely on streaming reasoning output, these kinds of fixes often resolve real stability concerns in production environments.

Update Overview

According to the GitHub release notes, @ai-sdk/xai@2.0.94 falls under Patch Changes, with the core fix coming from commit 2c72feb:

fix(deepseek): preserve reasoning streams across empty tool-call deltas

In plain terms, this fix addresses the problem of preserving reasoning streams across empty tool-call deltas. When a model produces an empty tool-call data chunk during streaming output, the previous behavior could cause the reasoning stream to be unexpectedly interrupted or lost. This update ensures the reasoning stream is fully preserved in this edge case.

The release was automatically published via GitHub Actions on September 11, with a verified GPG signature (key ID: B5690EEEBB952194). This automated release and signature verification process reflects Vercel AI's commitment to supply chain security best practices.

Why This Fix Matters

Although the fix is namespaced under deepseek, its appearance in the xai package release notes indicates that reasoning stream handling logic is shared across providers within the Vercel AI SDK. As models from DeepSeek, xAI (Grok), and others increasingly introduce "reasoning content" as a separate output channel distinct from the final answer, correctly parsing and preserving these reasoning tokens at the SDK layer becomes a critical engineering detail.

A "reasoning stream" refers to the model's chain-of-thought output produced before delivering its final response. This content is often interleaved with tool calls and returned incrementally as deltas during streaming. When a particular delta happens to contain empty tool-call data, a poorly handled parsing edge case can mistakenly signal that the entire reasoning stream has ended — causing subsequent content to be dropped. This fix specifically targets that boundary condition.

Practical Impact for Developers

For applications using the Vercel AI SDK to integrate xAI or DeepSeek models with streaming reasoning output, upgrading to 2.0.94 is recommended. In particular, projects that use both function calling (tool calls) and reasoning display features may have experienced intermittent missing reasoning content — an issue that should be resolved in this version.

Upgrading follows the standard npm package workflow:

npm install @ai-sdk/xai@2.0.94

As a patch release following Semantic Versioning (SemVer), it introduces no breaking changes and can theoretically be applied safely without any modifications to existing code.

What Frequent Iteration Tells Us About the AI SDK Ecosystem

Vercel's ai repository currently has over 27,000 Stars and 5,100 Forks on GitHub, with the main branch accumulating more than 4,000 commits since this version. This rapid iteration pace reflects how quickly the AI application development toolchain is evolving. Every seemingly small patch represents a real issue surfaced from actual usage.

For teams building AI-native applications, keeping up with SDK version updates and promptly applying fixes to core capabilities like reasoning streams and tool calls is essential for maintaining a stable product experience. The @ai-sdk/xai@2.0.94 update may be small, but it's a telling example of this ongoing refinement process.