deniz.in

Markets

Weather

Loading weather

· via Vercel blog

Vercel's AI Gateway adds Private Inference for self-hosted models and Seedance 2.5 Flash video

Vercel's AI Gateway can now route self-hosted and fine-tuned models through the same auth and observability as its catalog, and ByteDance's faster, cheaper Seedance 2.5 Flash video model has joined the lineup.

Vercel's AI Gateway adds Private Inference for self-hosted models and Seedance 2.5 Flash video

Vercel has rolled out two changes to AI Gateway, its routing layer for model APIs. Private Inference lets organizations connect their own model endpoints to the gateway and manage them like catalog models, while ByteDance's Seedance 2.5 Flash, a faster and cheaper cut of the company's video generation model, is now available to gateway users. Both announcements appeared in Vercel's changelog.

Self-hosted models join the gateway

According to Vercel, organizations running fine-tuned, self-hosted, or prerelease models have had to juggle separate credentials, bespoke client code, and their own logging to use those models alongside hosted ones. Private Inference collapses that split: a team's own endpoints now sit behind the gateway's shared authentication, SDKs, and APIs, and their traffic lands in the same observability tooling as everything else.

Format translation is handled by the gateway itself. Calls made through the Chat Completions, Responses, or Messages interfaces are converted to whatever shape the destination endpoint expects, which means swapping an application between a private model and a catalog model comes down to changing the model ID.

The feature also covers access control. Private models are visible only to the Vercel team that registered them and never appear in the public catalog; stored credentials are encrypted, and Zero Data Retention can be switched on per provider. Coding agents connected to the gateway, which Vercel lists as including Claude Code, Codex, and Cursor, can use private models like any other. Request logs, traces, and custom reports cover private traffic alongside catalog traffic, with charts splitting requests, token counts, and time to first token by private model. Administrators can also configure a catalog model as a fallback if a private endpoint fails.

Registering a private provider

Setup happens either on the dashboard's Private Inference page or through Vercel CLI 65.16.0 and newer. An admin registers an endpoint as a private provider by supplying its base URL, picking its API format, and adding credentials, then lists the models it serves. Applications call those models through IDs namespaced under private/, and the CLI example in Vercel's post wires up a self-hosted vLLM endpoint exposing a fine-tuned model. Supported model types, requirements, and billing details are covered in the documentation.

Seedance 2.5 Flash lands on the catalog

In the second post, Vercel confirmed that ByteDance's Seedance 2.5 Flash is available through AI Gateway. It is positioned as a speed- and cost-optimized version of Seedance 2.5, and Vercel says quality tracks the standard model across human realism, reference-based editing, and prompt adherence.

The model produces video with synchronized audio from text prompts, or from image, video, and audio references. Clips run four to 30 seconds at up to 720p and 24 frames per second, in landscape, square, or portrait aspect ratios. Existing videos can be edited or extended, and still images can be supplied to fix the first and last frames of a generation.

Developers invoke it as bytedance/seedance-2.5-flash with AI SDK 7. Reference material is passed as URLs through inputReferences and addressed in the prompt as [Image 1], [Video 1], and so on. Vercel also points to its model playground and a video generation quickstart for anyone evaluating the model.

Why it matters

Both updates push toward the same outcome: AI Gateway as the single control plane for every model an organization touches, wherever that model runs. Private Inference is the strategically heavier of the two. Enterprises with fine-tuned or open-weight models on their own hardware no longer need a parallel stack of keys, clients, and logging to combine them with commercial APIs, and they inherit the gateway's agent support, per-model metrics, and fallback routing in the process. Seedance 2.5 Flash extends the now-familiar pattern of flash-tier models to video: near-flagship output at lower cost and latency makes audio-synced generated video viable for a wider range of products. For teams already standardized on Vercel's tooling, neither change required new integration work, only configuration.

  • #vercel
  • #ai-gateway
  • #ai-inference
  • #video-generation
  • #bytedance

Related posts