---
type: analysis
slug: research-ling-3-0-flash-sante-s-free-window-has-ended-it-s-now-a-cheap-paid-api-with-no
title: "Ling 3.0 Flash Sante's free window has ended; it's now a cheap paid API with no published medical benchmarks"
description: "The health-tuned version of Ant Group's 124B-parameter Ling 3.0 Flash was free on Vercel AI Gateway from September 4 to October 4, 2026. The free ID no longer serves requests. The paid ID runs through a single provider (Novita). We found no published medical evaluations and no released weights."
published_at: 2026-10-23T19:00:00.000Z
topics: ["moe","benchmarks","reasoning","research"]
source_urls: ["https://vercel.com/changelog/ling-3-0-flash-sante-is-now-available-on-ai-gateway-for-free","https://vercel.com/ai-gateway/models/ling-3.0-flash-sante","https://huggingface.co/inclusionAI/Ling-3.0-flash","https://vercel.com/changelog/ling-3-0-flash-is-now-available-on-ai-gateway","https://vercel.com/ai-gateway/models/ling-3.0-flash-fin-free","https://huggingface.co/inclusionAI","https://huggingface.co/inclusionAI/Ling-3.0-flash-Sante","https://www.modelscope.cn/models/inclusionAI/Ling-3.0-Flash-Sante"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On September 4, 2026, Vercel added Ling 3.0 Flash Sante from inclusionAI (Ant Group's Ling team) to its AI Gateway. It was free through October 4. Vercel calls it the healthcare-focused version of Ling 3.0 Flash, built for medical reasoning, research, evidence retrieval and multi-step healthcare tasks. It has a 256K context window and function calling, and keeps the base model's general reasoning, coding and agentic skills. The X post that flagged the story is Vercel's own promotion of that changelog.

The release matters less for this one model than for how it was shipped: one open-weight sparse MoE backbone, cut into domain-specific variants (health, finance, vision), each launched through a gateway with a free window. Builders need to know what remains after the window closes, what the model actually is, and what evidence supports its use in a regulated domain like health.

## What we checked

- **The cluster may reveal a broader market or architecture transition that cannot be responsibly reduced to a single release note.** partly confirmed. There is a pattern, but it is a distribution pattern more than an architecture shift. Vercel has launched Ling 3.0 Flash variants with time-limited free IDs: the base model was free through August 3, and Sante through October 4. A finance version (Fin) has a '-free' listing too. All of these share one hybrid KDA/MLA MoE backbone and differ by domain fine-tuning. The release note alone doesn't show a market-wide transition. [4][1][5][3]
- **(Added) Are open weights for the Sante variant publicly available, like the MIT-licensed base model?** unresolved. The base Ling-3.0-flash card on Hugging Face is MIT-licensed. inclusionAI's Hugging Face organization page listed no Sante repository, and a guessed Sante repo URL returned HTTP 401, which usually means the repo is missing or private. A ModelScope page titled Ling-3.0-Flash-Sante exists, but we could not read its contents. Treat Sante as API-only until a model card appears. [6][3][7][8]
- **(Added) Is the 'medical reasoning and evidence-based retrieval' positioning backed by published health benchmarks?** unresolved. Neither Vercel's changelog nor its model page gives any medical benchmark scores. We found no inclusionAI primary source with health evaluations, so the domain claims remain vendor positioning only. [1][2]

## What we found

- The free period is over. Vercel's changelog gave two IDs: `inclusionai/ling-3.0-flash-sante-free` stops serving requests after October 4, and `inclusionai/ling-3.0-flash-sante` continues at standard rates. Free requests still showed up in usage and traces at $0 cost. Any integration still pointing at the -free ID is now failing. [1]
- Vercel's model page lists the paid ID at $0.075 per 1M input tokens and $0.22 per 1M output tokens. Context is 256K and maximum output is 32K. The page names one provider, Novita, so there is no failover across providers within the gateway. The same page describes a MoE with about 124B total and about 5.1B active parameters per token. [2]
- The base model's Hugging Face card describes a hybrid-linear MoE. Kimi Delta Attention and gated Multi-head Latent Attention layers alternate 5:1 (35 KDA layers, 7 MLA layers). There are 512 routed experts plus 1 shared, with 8 active. Training ran on an 8K→32K→256K context schedule, and the card shows serving at 262,144 tokens with YaRN. The license is MIT. Sante presumably inherits this backbone, but we found no Sante-specific technical card that confirms it. [3]
- The free launch follows the base model's. Vercel's July 23, 2026 changelog for Ling 3.0 Flash offered `inclusionai/ling-3.0-flash-free` through August 3, with the same 124B/5.1B, 256K and thinking/non-thinking specs. Vercel also lists a finance variant, Ling 3.0 Flash Fin (Free), served by Novita, but that page gives no dates for its free window. [4][5]
- We could not confirm open weights for Sante. inclusionAI's Hugging Face organization page showed Ling 3.0 UBTree models and Ling-3.1-flash, but no Sante repository, and a direct Sante repo URL returned HTTP 401. A ModelScope page titled Ling-3.0-Flash-Sante exists, but its contents could not be read. We could not open an inclusionAI primary announcement or any medical benchmark table. [6][7][8]

## What it means

If you tried Sante during the free window, check your config first. Calls to the -free ID stopped after October 4. Move to `inclusionai/ling-3.0-flash-sante` (currently $0.075/$0.22 per million tokens) or remove it. Because Vercel lists a single provider, Novita, put your own fallback model behind it. At about 5.1B active parameters, cost and latency should be low, so it is a reasonable candidate for high-volume, low-stakes work such as triaging literature, summarizing guidelines, or drafting retrieval-augmented answers that a clinician reviews. Vercel has published no health benchmarks, and we found no published evaluations or weights from inclusionAI. Run it on your own medical eval set and compare it against the general Ling 3.0 Flash base. A domain fine-tune doesn't automatically beat its base model, and nothing here supports clinical decision-making without human oversight.

The bigger lesson is about distribution. Labs are taking one cheap open-weight sparse MoE backbone and shipping vertical fine-tunes through gateways with short free trials. The free window is a way to get people trying the model, not a commitment to keep it free. Pin model IDs, watch changelogs for end dates, and don't assume a vertical variant inherits the base model's MIT license or open weights. Until there is a Sante model card, treat it as an API-only product.

## Still open

- Has inclusionAI published any medical benchmark results (and against which baselines) for Sante, and does it actually outperform base Ling 3.0 Flash on health tasks?
- Will Sante weights be released under MIT like the base model, and what does the ModelScope listing actually contain?
- What data and safety tuning went into the health fine-tune, and has any third party audited it?
- Will Vercel AI Gateway add providers beyond Novita for the paid Sante ID?

## Sources

1. [Ling 3.0 Flash Sante is now available on AI Gateway for free (Vercel changelog)](https://vercel.com/changelog/ling-3-0-flash-sante-is-now-available-on-ai-gateway-for-free) (primary)
2. [Ling 3.0 Flash Sante – Vercel AI Gateway model page](https://vercel.com/ai-gateway/models/ling-3.0-flash-sante) (primary)
3. [inclusionAI/Ling-3.0-flash model card (Hugging Face)](https://huggingface.co/inclusionAI/Ling-3.0-flash) (primary)
4. [Ling 3.0 Flash is now available on AI Gateway (Vercel changelog)](https://vercel.com/changelog/ling-3-0-flash-is-now-available-on-ai-gateway) (primary)
5. [Ling 3.0 Flash Fin (Free) – Vercel AI Gateway model page](https://vercel.com/ai-gateway/models/ling-3.0-flash-fin-free) (primary)
6. [inclusionAI organization page (Hugging Face)](https://huggingface.co/inclusionAI) (primary)
7. [Attempted Sante repo URL (returned HTTP 401)](https://huggingface.co/inclusionAI/Ling-3.0-flash-Sante) (primary)
8. [Ling-3.0-Flash-Sante – ModelScope listing](https://www.modelscope.cn/models/inclusionAI/Ling-3.0-Flash-Sante) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Moe](https://newruntime.com/topics/moe.md): Explore the moe topic hub.
- topic hub: [Benchmarks](https://newruntime.com/topics/benchmarks.md): Explore the benchmarks topic hub.
- topic hub: [Reasoning](https://newruntime.com/topics/reasoning.md): Explore the reasoning topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)