Ling 3.0 Flash Sante is now available on AI Gateway for free

What shipped

inclusionAI's Ling 3.0 Flash Sante is now available on Vercel's AI Gateway at no cost through October 4. The model is a health and medicine-focused variant of Ling 3.0 Flash, built on a Mixture-of-Experts architecture with 124 billion total parameters and approximately 5.1 billion active per token, a 256,000-token context window, and native function calling support.

This piece breaks down what shipped, how the model is structured and who benefits from it, how to access it through AI Gateway today, where friction points might arise, and what developments are worth tracking as the free access window closes. It is aimed at developers building health-adjacent products or experimenting with large-context, domain-specialized inference on managed infrastructure.

inclusionAI released Ling 3.0 Flash Sante as a health and medicine-focused member of the Ling 3.0 Flash family, and Vercel has made it available through AI Gateway with zero cost through October 4. The model uses a Mixture-of-Experts design, meaning only a subset of its total weights activate for any given token — roughly 5.1 billion out of 124 billion — keeping inference efficient while preserving the representational depth of a much larger dense model. The 256,000-token context window is large enough to accommodate lengthy clinical documents, multi-turn diagnostic conversations, or extended research literature without chunking.

What improved

The combination of domain focus, sparse activation, and a long context window is the core of what makes this release distinct from a general-purpose addition to AI Gateway's model catalog. Vercel is offering access through its standard gateway interface, meaning existing projects that already route requests through AI Gateway can add Ling 3.0 Flash Sante without rebuilding their authentication or observability setup. Function calling is included, which is significant for builders who need the model to trigger structured retrieval or downstream API calls rather than just generate text.

Developers who route requests through Vercel AI Gateway now have a domain-tuned model available alongside general-purpose options, without any additional subscription or per-token fee until October 4. For teams experimenting with health-related use cases — clinical note summarization, patient-facing Q&A, medical coding assistance, or literature review — this removes the cost barrier that typically makes evaluation cycles slow. The 124B parameter scale and 5.1B active parameter profile mean the model occupies a different inference cost bracket than dense models of equivalent total size, which matters when projecting what post-October usage might look like if teams decide to continue.

What you gain

Function calling support means the model can participate in agentic workflows rather than operating purely as a text generator. Builders constructing pipelines that need to retrieve structured health data, call external APIs, or chain tool outputs into a longer reasoning loop can wire Ling 3.0 Flash Sante into those flows directly. The 256,000-token window means the model can hold a full conversation history plus retrieved context simultaneously, which reduces the complexity of context management code that teams would otherwise need to write and maintain.

Access runs through Vercel AI Gateway, so the starting point is the AI Gateway dashboard in your Vercel account. Ling 3.0 Flash Sante appears in the model catalog; selecting it generates the endpoint configuration your application needs to start sending requests. If you are already using AI Gateway with another model, updating the model identifier in your request payload is the main change — no new credentials or gateway instances are required.

For teams using the Vercel AI SDK, the gateway abstraction means switching models involves changing the model string and verifying that your prompt structure aligns with what Ling 3.0 Flash Sante expects. Function calling configuration follows the same schema AI Gateway exposes for other models that support it, so existing tool definitions should transfer with minimal adjustment. Confirming that your token budget fits within the 256,000-token limit is worth doing early, since requests that exceed context limits fail at the API layer rather than silently truncating.

How to get it

The free access window closes on October 4, and inclusionAI and Vercel have not announced what the pricing structure will look like after that date. Teams that build production workflows on this model during the free period should treat October 4 as a hard planning deadline: either confirm post-window pricing before committing, or build a model-swap path so the application can fall back to an alternative without a service interruption. Assuming continued free access beyond that date is a risk worth flagging during architecture review.

Domain-specific models sometimes exhibit degraded performance on general-purpose tasks relative to their specialized domain. Builders who route mixed traffic — some health-related, some not — through a single model endpoint should test Ling 3.0 Flash Sante against their full request distribution, not just the medical subset. The 5.1 billion active parameters per token figure describes the model's efficiency profile, but actual latency and throughput under your specific load pattern will depend on AI Gateway's serving configuration, which may differ across regions.

What to watch next

The October 4 deadline is the most concrete near-term signal to track. Any pricing announcement from inclusionAI or Vercel before that date will determine whether the model becomes a viable long-term component of health-adjacent stacks or primarily a low-cost evaluation window. Teams should also watch for updates to the Ling 3.0 Flash family more broadly, since Sante's release suggests inclusionAI is pursuing domain-specific variants alongside general-purpose improvements.

On the infrastructure side, AI Gateway's expanding model catalog is worth monitoring independently of any single model. As more specialized and non-English-first models appear through the same interface, the gateway layer becomes a more important abstraction for teams that want to swap models without rewriting routing or observability logic. Ling 3.0 Flash Sante's addition is one data point in that pattern, and how it performs in production during the free window will inform how seriously teams engage with domain-tuned alternatives going forward.