# Ling-3.0-flash-Sante

> Source: https://aiwiki.ai/wiki/ling_3_0_flash_sante
> Updated: 2026-09-05
> Fact-checked: 2026-09-05
> Categories: AI Models, Chinese AI, Healthcare AI, Mixture of Experts
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Ling-3.0-flash-Sante." aiwiki.ai, 5 Sept 2026. https://aiwiki.ai/wiki/ling_3_0_flash_sante
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Ling-3.0-flash-Sante** is a health and medicine variant of [Ling-3.0-flash](https://aiwiki.ai/wiki/ling_3_0_flash), the 124-billion-parameter [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) language model from [inclusionAI](https://aiwiki.ai/wiki/inclusionai), the open-source AI initiative of [Ant Group](https://aiwiki.ai/wiki/ant_group). The Ant Ling team announced it on September 4, 2026, describing it as "an MoE model enhanced for health and medicine, built on Ling-3.0-flash" and explaining that the name is "Inspired by the French word 'santé,' meaning 'health'".[1] It keeps the base model's 124B total and about 5.1B active parameters per token and its 256K-token context window.[2][4] As of September 5, 2026, it is available only as a hosted API, through [OpenRouter](https://aiwiki.ai/wiki/openrouter) (free tier), [Vercel](https://aiwiki.ai/wiki/vercel) AI Gateway (free through October 4, 2026), and Novita AI; no weights have been published on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) or [ModelScope](https://aiwiki.ai/wiki/modelscope), and Ant's own developer documentation does not yet list the model.[2][4][6][7][8][9]

The launch post carried a single benchmark chart comparing Sante with nine other models on twelve evaluations, seven of them medical. All of those numbers are Ant-reported and come from an image; this article transcribes them with the chart's own caveats. Ant has not published a model card, a technical report, training details, or a safety evaluation for the model.[1]

## Announcement

The announcement came from the @AntLingAGI account on X at 19:14 UTC on September 4, 2026. The full text, apart from a closing toast, reads: "Today, we're introducing Ling-3.0-flash-Sante" followed by the description "an MoE model enhanced for health and medicine, built on Ling-3.0-flash." The second paragraph gives the name and the intended use: "Inspired by the French word 'santé,' meaning 'health,' Sante is built for real-world healthcare tasks spanning medical reasoning, professional healthcare tasks, deep research, and evidence-based retrieval." The third makes the performance claim: "Sante shows competitive results across MedXpertQA-Text, DiagnosisArena-MCQ, AFUMED-Drug, HealthBench Professional, and BrowseComp, with leading performance among open-source models and results competitive with flagship models." The post ends "Here's to better health" and a "Santé!", and points to a thread that was not retrievable at the time of writing.[1]

Sante was the second Ling 3.0 variant announced that day. Seventy minutes earlier, at 18:04 UTC, the same account had posted that Ant was "releasing Ling-3.0-flash-VL, built on Ling-3.0-flash with visual understanding and visual agent capabilities", a model it said "performs well across visual perception, STEM reasoning, document intelligence, multimodal agent tasks, frontend coding, and medical report interpretation".[30] The OpenRouter listing for Sante was created at 18:19 UTC, before the Sante post itself, and Vercel's changelog entry and Novita's blog post are both dated September 4.[3][4][11]

## What the model is

Sante is a domain-adapted version of Ling-3.0-flash rather than a new architecture. The hosting platforms describe it in near-identical terms. OpenRouter's listing calls it "a health and medicine-focused mixture-of-experts model from InclusionAI, built on Ling 3.0 Flash with 5.1B active parameters out of 124B total", "designed for medical knowledge reasoning, clinical safety, evidence-based retrieval, and long-horizon medical tasks, while retaining general capabilities in reasoning, coding, and agentic tasks".[2] Vercel's changelog says it "is a Mixture-of-Experts model with 124B total parameters and about 5.1B active per token, a 256K token context window, and function calling", "built for medical reasoning, professional healthcare tasks, deep research, evidence-based retrieval, and multi-step medical workflows".[4] Novita's model page uses the same list of strengths, adding that the model "retain[s] strong general capabilities in reasoning, coding, and agentic tasks".[12]

Nothing published so far says how the adaptation was done. Ant's finance variant, Ling-3.0-flash-Fin, was produced "through continued training on high-quality financial data" with "leading financial institutions and domain experts", according to its model card; whether Sante followed the same recipe with medical data has not been stated.[28] No training-data description, evaluation methodology beyond the chart notes, or safety assessment has been released.

The base model's design is covered on the Ling-3.0-flash page. In brief: 124B total parameters with about 5.1B active per token; 42 transformer layers alternating Kimi Delta Attention, a form of [linear attention](https://aiwiki.ai/wiki/linear_attention), with gated [multi-head latent attention](https://aiwiki.ai/wiki/multi-head_latent_attention) at a 5:1 ratio; 512 routed experts plus one shared, with 8 routed experts active per token (a 1/64 activation ratio); a native 256K context that Ant says can scale to 1M tokens; a hybrid thinking mode that can be switched off per request; and [MIT-licensed](https://aiwiki.ai/wiki/mit_license) weights released on August 2, 2026.[13][14] Ant's developer documentation lists the base model's peak throughput at 1,000 tokens per second with a time to first token under 100 milliseconds; those are the company's figures for the base model, not measurements of Sante.[10]

## Availability

As of September 5, 2026, Sante can be called through three hosted routes, all of which resolve to the same provider. OpenRouter's page says the model "is hosted by one provider", and its endpoints API names that provider as Novita; Vercel's model page likewise lists Novita AI as the sole provider.[2][3][5] Novita's blog states that the model "was added to Novita's model catalog on September 3, 2026", a day before the public announcement, and that its listing was "marked as time limited free" on September 4.[11]

| Platform | Model id | Price (Sep 5, 2026) | Context | Max output | Features listed | Listed date |
| --- | --- | --- | --- | --- | --- | --- |
| [OpenRouter](https://aiwiki.ai/wiki/openrouter) | `inclusionai/ling-3.0-flash-sante:free` | Free (rate-limited free tier) | 262,144 tokens | 32,768 tokens | Tools and `tool_choice`; reasoning on by default; no `response_format` | Created Sep 4, 2026, 18:19 UTC; page says "released on September 4, 2026" |
| [Vercel](https://aiwiki.ai/wiki/vercel) AI Gateway | `inclusionai/ling-3.0-flash-sante` (bills after the offer) and `inclusionai/ling-3.0-flash-sante-free` (stops serving after the offer) | Free through October 4, 2026 | 256K | 32K | [Function calling](https://aiwiki.ai/wiki/function_calling); provider Novita AI | Sep 4, 2026 |
| Novita AI (serverless API) | `inclusionai/ling-3.0-flash-sante` | $0 per million input and output tokens, "time limited free" | 262,144 tokens | 32,768 tokens | Reasoning, function calling; OpenAI-compatible and Anthropic-compatible endpoints; text in, text out | Catalog Sep 3, 2026 (per Novita) |
| [Hugging Face](https://aiwiki.ai/wiki/hugging_face) | none | not applicable | not applicable | not applicable | No repository named Ling-3.0-flash-Sante under inclusionAI as of Sep 5, 2026 | not applicable |
| [ModelScope](https://aiwiki.ai/wiki/modelscope) | none | not applicable | not applicable | not applicable | No inclusionAI/Ling-3.0-flash-Sante entry as of Sep 5, 2026 (the API returns "record not found") | not applicable |
| Ant Ling API (developer.ant-ling.com) | none listed | not applicable | not applicable | not applicable | Changelog (last updated Sep 4, 2026) records Ling-3.0-flash-VL but no Sante entry; pricing page has no Sante row | not applicable |

OpenRouter's models API reports the free listing with a 262,144-token context, a 32,768-token completion limit, zero prompt and completion pricing, and support for `tools`, `tool_choice`, `reasoning`, `logprobs`, `seed`, and the usual sampling parameters; `response_format` is absent, and the page states that "JSON output is not enforced".[2][3] The same API gives the canonical slug `inclusionai/ling-3.0-flash-sante-20260904` and no linked Hugging Face repository.[3]

Vercel's changelog, by Zachary Chen and Jerilyn Zheng, explains the two ids: "The standard model ID, inclusionai/ling-3.0-flash-sante, is free through October 4 and begins billing when the offer ends. The free model ID, inclusionai/ling-3.0-flash-sante-free, stops serving when the offer ends instead of billing." Free requests "still appear in your spend dashboard and carry a trace". The post says AI Gateway "reflects provider pricing with no markup"; the price that will apply after October 4 is not stated on the changelog or the model page, whose provider table shows "Free" for input and output and the note "Promotional pricing ends on October 4, 2026".[4][5]

The pattern repeats the launch of the base model, which Ant made free on OpenRouter and Vercel AI Gateway for a limited period before publishing weights, and of the finance variant, which OpenRouter listed on August 27, 2026 with weights following on September 3.[14][29][28] Whether Sante will follow the same path to open weights has not been announced; the Hugging Face listing for the base model is MIT-licensed, but no license has been stated for Sante.[6][13]

## Reported benchmarks

Ant's only published evaluation of Sante is the chart attached to the announcement post. It compares Sante with nine models on twelve benchmarks: seven medical evaluations (two of them internal to Ant), two Ant-labelled safety and ethics evaluations, and three general agentic-search evaluations. The table below transcribes the bar labels as they appear in the image. Values are Ant-reported, from a chart in the announcement; none has been independently reproduced. Model names and parameter counts are as the chart's legend gives them; "max" and "high" in the legend denote the reasoning-effort setting Ant says it used for that model.[1]

| Benchmark | Ling-3.0-flash-Sante (124B A5.1B) | [DeepSeek-V4-Flash](https://aiwiki.ai/wiki/deepseek_v4_flash)-0731 (max) (284B A13B) | [GLM-5.3-Flash](https://aiwiki.ai/wiki/glm_5_3_flash) (max) (320B A18B) | [MiniMax-M2.7](https://aiwiki.ai/wiki/minimax_m2_7) (229.9B A9.8B) | Step-3.7-Flash (high) (198B A11B) | [Nemotron-3-Super](https://aiwiki.ai/wiki/nemotron_3) (120B A12B) | [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) (max) (2.8T A104B) | [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8) (max) | [GPT-5.6](https://aiwiki.ai/wiki/gpt_5_6) Sol (max) | [Gemini 3.6 Flash](https://aiwiki.ai/wiki/gemini_3_6_flash) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| MedXpertQA-Text | 53.9 | 41.5 | 38.3 | 35.4 | 35.6 | 36.6 | 53.5 | 50.5 | 60.2 | 62.4 |
| [HealthBench](https://aiwiki.ai/wiki/healthbench) Professional | 45.7 | 33.9 | 49.1 | 26.0 | 23.4 | 21.9 | 49.6 | 49.4 | 60.8 | 45.8 |
| DiagnosisArena-MCQ | 83.8 | 72.8 | 74.2 | 58.5 | 56.3 | 54.0 | 78.4 | 74.9 | 81.9 | 76.2 |
| ReDisQA | 91.3 | 91.0 | 91.4 | 87.8 | 86.8 | 87.8 | 90.8 | 90.8 | 91.6 | 92.4 |
| [MedQA](https://aiwiki.ai/wiki/medqa)-USMLE | 93.5 | 94.5 | 95.2 | 90.5 | 91.4 | 91.9 | 95.0 | 94.3 | 96.7 | 94.7 |
| MedMCQA | 81.6 | 82.8 | 85.1 | 77.4 | 78.7 | 77.7 | 86.9 | 85.3 | 87.6 | 87.0 |
| MedEthicAlign | 82.1 | 81.3 | 81.0 | 76.7 | 75.6 | 85.6 | 86.3 | 80.9 | 73.6 | 81.5 |
| AFUMED-Drug (internal) | 89.6 | 86.3 | 82.3 | 83.2 | 82.7 | 83.5 | 86.8 | 88.4 | 92.0 | 84.4 |
| AFUSAFE-MedSCE (internal) | 78.6 | 81.6 | 81.5 | 72.3 | 78.7 | 78.6 | 81.0 | 81.7 | 79.9 | 81.0 |
| [BrowseComp](https://aiwiki.ai/wiki/browsecomp), single agent | 73.9 | 79.5 | not shown | 77.8 | 75.8 | 31.3 | 91.2 | 84.3 | 90.4 | not shown |
| BrowseComp, multi-agent | 86.9 | not shown | not shown | not shown | not shown | not shown | not shown | 88.5 | not shown | not shown |
| DeepSearchQA | 86.8 | 89.3 | not shown | 83.2 | 92.8 | 71.9 | 95.0 | 93.1 | not shown | 83.8 |
| [HLE](https://aiwiki.ai/wiki/humanity_s_last_exam) with tools | 53.2 | 51.5 | 55.3 | not shown | 47.2 | 22.8 | 56.0 | 57.9 | not shown | 33.1 |

"Not shown" means the chart has no bar for that model on that panel; Ant gives no reason for the gaps. On the BrowseComp panel, solid bars are labelled "Single Agent with Context" and hatched bars "Multi-Agent"; only Sante and Claude Opus 4.8 have both.[1]

Read against its own comparators, the chart supports a narrower statement than the post's "leading performance among open-source models". Among the seven open-weight models in the legend, Sante has the highest bar on three of the twelve panels: MedXpertQA-Text (53.9 against Kimi K3's 53.5), DiagnosisArena-MCQ (83.8 against 78.4), and Ant's internal AFUMED-Drug (89.6 against 86.8). On the other nine it is behind at least one open-weight model, most often Kimi K3, which the legend lists at 2.8T total parameters, or GLM-5.3-Flash; on HealthBench Professional it trails both (45.7 against 49.6 and 49.1).[1] Sante has the smallest active-parameter count among the models whose sizes the legend gives. Across all ten models, Sante has the highest bar on DiagnosisArena-MCQ (83.8 against GPT-5.6 Sol's 81.9); GPT-5.6 Sol has the highest bar on HealthBench Professional, MedQA-USMLE, MedMCQA, and AFUMED-Drug; Gemini 3.6 Flash on MedXpertQA-Text and ReDisQA; Claude Opus 4.8 on AFUSAFE-MedSCE and HLE with tools; and Kimi K3 on MedEthicAlign, BrowseComp, and DeepSearchQA. On MedEthicAlign, GPT-5.6 Sol's 73.6 is the lowest bar of the panel.[1]

### Evaluation settings stated on the chart

The chart's notes, reproduced here because they change how the numbers should be read, are Ant's own:[1]

| Item | Ant's note |
| --- | --- |
| Sampling | "Unless otherwise specified, Ling-3.0-flash-Sante is evaluated with temperature=0.6 and top_p=0.95 across all benchmarks." |
| Reasoning effort | "The reasoning effort was set to the highest available tier for each model." |
| AFUMED-Drug | "a benchmark that emphasizes medication safety. It primarily evaluates model accuracy in identifying medication contraindications." Marked "(internal)". |
| MedEthicAlign | "a high-quality subset of 1,000 samples is used, covering compliance and ethical reasoning in high-risk Chinese medical scenarios." |
| AFUSAFE-MedSCE | "an in-house benchmark covering medical ethics and safety, built from real-world risk interactions and recent regulatory updates." Marked "(internal)". |
| MedMCQA | "the standard validation set is used for evaluation." |
| BrowseComp | Temperature 1.0, top_p 0.95, 256K context window, maximum 1,000 turns. "The single-agent setting uses a discard-all context-management strategy, triggered when context usage reaches 40% of the maximum window size. For the multi-agent setting, we use an in-house harness capable of dispatching up to 64 subagents to complete each task." |
| HLE with tools and DeepSearchQA | Temperature 1.0, top_p 0.95, 256K context window. "No context-management strategy is applied." |

Three consequences follow. First, two of the twelve panels, AFUMED-Drug and AFUSAFE-MedSCE, are Ant's own unpublished benchmarks, so nobody outside Ant can check them. Second, the BrowseComp setup differs from the one Ant used for the base model, whose card describes a single-agent run that summarizes and resumes at a 64K-token threshold and a multi-agent harness with 128K and 64K context windows; Sante's chart instead uses a discard-all strategy at 40 percent of a 256K window and a harness of up to 64 subagents, so the two models' BrowseComp figures are not like-for-like.[13] Third, the comparators were run by Ant in Ant's harness at the highest reasoning effort available, and the chart does not say which API versions or dates were used.

### The public benchmarks

| Benchmark | What it measures | Source |
| --- | --- | --- |
| MedXpertQA-Text | The text subset of MedXpertQA, a benchmark of 4,460 expert-level medical questions across 17 specialties and 11 body systems, built with "rigorous filtering and augmentation to address the insufficient difficulty of existing benchmarks like MedQA" and including specialty board questions. The MM subset adds images; the chart uses Text only. | Zuo et al., Tsinghua, January 2025 [15] |
| HealthBench Professional | OpenAI's April 2026 companion to HealthBench: "an open benchmark for evaluating large language models on real tasks that clinicians bring to ChatGPT", organized around care consult, writing and documentation, and medical research, with physician-authored conversations scored by rubrics "written and iteratively adjudicated by three or more physicians". Examples were selected for difficulty against OpenAI's frontier models. The original HealthBench (May 2025) has 5,000 multi-turn conversations graded with 48,562 physician-written rubric criteria. | Soskin Hicks et al., OpenAI, April 2026 [17]; Arora et al., OpenAI, May 2025 [16] |
| DiagnosisArena-MCQ | DiagnosisArena is 1,113 segmented patient cases with diagnoses from case reports in 10 medical journals, spanning 28 specialties, designed as an open-ended diagnostic task. The paper defines DiagnosisArena-MCQ as a conversion "into the multiple-choice format based on model-generated diagnoses" and notes that "multiple-choice formats inherently reduce task difficulty". | Zhu et al., May 2025 (v5 August 2026) [18] |
| ReDisQA | ReDis-QA, a rare-disease question-answering set of 1,360 question-answer pairs covering 205 rare diseases, with per-question metadata, released with the ReCOP corpus drawn from the National Organization for Rare Disorders database. | Wang et al., August 2024 [19] |
| MedQA-USMLE | The English portion of MedQA, "the first free-form multiple-choice OpenQA dataset for solving medical problems", collected from professional medical board exams; the English set has 12,723 questions from the United States Medical Licensing Examination. | Jin et al., MIT, September 2020 [20] |
| MedMCQA | "More than 194k" multiple-choice questions from the Indian AIIMS and NEET PG medical entrance exams, covering 2.4k healthcare topics and 21 subjects; Ant reports the standard validation split. | Pal et al., March 2022 [21] |
| MedEthicAlign | A GitHub repository of this name, maintained by the X-LANCE lab, hosts the MedES benchmark from an AAAI 2026 paper: a Chinese medical-ethics benchmark "constructed from 260 authoritative Chinese medical, ethical, and legal sources" with 1,179 objective and 4,897 subjective questions. Ant's chart describes its MedEthicAlign as a 1,000-sample subset covering "high-risk Chinese medical scenarios" but does not say which release or split it used. | Jin et al., X-LANCE, AAAI 2026 [22] |
| BrowseComp | OpenAI's 1,266-question benchmark for browsing agents, with questions "that require persistently navigating the internet in search of hard-to-find, entangled information" and short, verifiable answers. | Wei et al., OpenAI, April 2025 [23] |
| DeepSearchQA | A 900-prompt benchmark from Google DeepMind for "difficult multi-step information-seeking tasks across 17 different fields" whose tasks ask for exhaustive answer lists and are structured as causal chains. | Gupta et al., Google DeepMind, January 2026 [24] |
| HLE with tools | Humanity's Last Exam, 2,500 closed-ended expert questions across dozens of subjects, written by subject-matter experts to sit "at the frontier of human knowledge"; Ant's run allows tool use. | Phan et al., Center for AI Safety and Scale AI, January 2025 [25] |

## Domain variants of Ling-3.0-flash

Sante is the third specialized version of the flash model that Ant has announced since the base weights shipped, and its rollout follows the pattern set by the finance variant.[14]

| Variant | Announced | Hosted API | Open weights | Status of evaluation |
| --- | --- | --- | --- | --- |
| Ling-3.0-flash-Fin (finance) | August 27, 2026 (X post) | OpenRouter listing created August 27, 2026; paid tier and free tier | Hugging Face `inclusionAI/Ling-3.0-flash-Fin`, MIT, created September 3, 2026 | Ant named FinFIRST, FinSearchComp Verified, FinCRAFT, FinanceAgent, APEX-Agents, SpreadsheetBench, and tau3-Banking; results published as an image |
| Ling-3.0-flash-VL (vision) | September 4, 2026, 18:04 UTC (X post) | Ant Ling platform per the developer changelog of September 4, 2026 ("Launched the multimodal Ling-3.0-flash-VL model"); not on OpenRouter | None found on Hugging Face or ModelScope as of September 5, 2026 | Benchmark image in the post |
| Ling-3.0-flash-Sante (health) | September 4, 2026, 19:14 UTC (X post) | OpenRouter, Vercel AI Gateway, Novita AI | None as of September 5, 2026 | Benchmark image in the post |

For the finance model, Ant wrote on August 27 that it "will open-source the model weights next week", TechNode reported on August 28 that the weights were "scheduled to be released next week" alongside "a one-month free API period through OpenRouter for finance professionals and developers", and the Hugging Face repository appeared on September 3.[26][27][28] The Fin model card carries a use disclaimer, stating that "valuation results, and investment conclusions require professional review and do not constitute investment advice", and recommends a sampling temperature of 1.0 rather than the base model's 0.6.[28] Sante's chart, by contrast, says it was evaluated at temperature 0.6.[1]

Ant's developer documentation describes "Domain-adapted models for specialized scenarios such as healthcare and finance, with superior performance on domain knowledge-intensive tasks" as one pillar of its Ling ecosystem, which places Sante within a planned line rather than a one-off.[10]

## Reception

No press coverage of Sante had appeared by September 5, 2026. A Bing News search for the model name returned no articles, and Hacker News searches for "Ling-3.0-flash-Sante" and "Ling 3.0 Sante" returned no submissions.[31] The only third-party writing found was from hosting platforms: Vercel's changelog, Novita's blog, and automatically generated model-catalog pages that copy OpenRouter's description. Novita's post is notable mainly for its framing: it says the model "is best treated as a developer API for research, retrieval, summarization, and workflow assistance, not as a medical device or a replacement for qualified clinical judgment", notes that the Novita page "does not provide a public benchmark table that would justify a clinical-accuracy ranking", and advises users not to "use a generated response as a diagnosis, treatment instruction, or independent clinical decision".[11] The announcement post itself had 252 likes, 47 reposts, and 27 replies about 11 hours after publication.[1]

## Limitations and unpublished details

Several things a reader would normally expect for a medical model were not available as of September 5, 2026:

| Item | Status |
| --- | --- |
| Weights | Not released; no repository on Hugging Face or ModelScope.[6][7] |
| License | Not stated for Sante. The base model and the Fin variant are MIT-licensed.[13][28] |
| Model card or technical report | None. The only first-party material is the X post and its chart.[1] |
| Training method and data | Not described. The Fin variant's card says that model used continued training on financial data; nothing equivalent exists for Sante.[28] |
| Safety evaluation | Ant's chart includes two ethics and safety panels (MedEthicAlign and the internal AFUSAFE-MedSCE), but no red-teaming, refusal, or clinical-harm evaluation has been published.[1] |
| Medical-use disclaimer | None on the OpenRouter or Vercel pages. Novita's blog states the model is not a medical device and that "Do not use a generated response as a diagnosis, treatment instruction, or independent clinical decision".[2][4][11] |
| Structured output | OpenRouter's endpoint does not support `response_format`; Novita's blog notes structured outputs are not among the listed features.[2][11] |
| Modality | Text in, text out. Medical images are out of scope; Ant's separate VL variant claims "medical report interpretation".[3][11][30] |
| Pricing after the free period | Not published. Vercel says the standard id "begins billing when the offer ends" on October 4, 2026; Novita labels its $0 price "time limited free".[4][11] |
| Regional or language coverage | Not stated. Two of the evaluations Ant chose (MedEthicAlign and AFUSAFE-MedSCE) are described in terms of Chinese medical scenarios and regulation.[1] |

The benchmark picture has its own gaps. Two panels are internal, three panels omit some comparators without explanation, and the two general-search panels (BrowseComp and DeepSearchQA) were run in Ant's own harness. The medical panels on which Sante leads the open-weight group are the ones with the lowest scores overall (MedXpertQA-Text and DiagnosisArena-MCQ), while on the licensing-exam sets (MedQA-USMLE, MedMCQA) it sits below the larger open models by one to about five points.[1] Until a model card, weights, or an independent evaluation appears, the chart is the only evidence for what the model does.

## References

1. [Ant Ling (@AntLingAGI) post introducing Ling-3.0-flash-Sante, with benchmark chart](https://x.com/AntLingAGI/status/2095953758148853892) - X, September 4, 2026, 19:14 UTC (chart image: pbs.twimg.com/media/HRZFtrJbQAAA06-.jpg; engagement counts as of September 5, 2026).
2. [inclusionAI: Ling 3.0 Flash Sante (free)](https://openrouter.ai/inclusionai/ling-3.0-flash-sante:free) - OpenRouter, accessed September 5, 2026.
3. [OpenRouter API, endpoints for inclusionai/ling-3.0-flash-sante:free](https://openrouter.ai/api/v1/models/inclusionai/ling-3.0-flash-sante:free/endpoints) - OpenRouter, accessed September 5, 2026 (creation timestamp, provider, context, completion limit, supported parameters); model list at [openrouter.ai/api/v1/models](https://openrouter.ai/api/v1/models).
4. [Ling 3.0 Flash Sante is now available on AI Gateway for free](https://vercel.com/changelog/ling-3-0-flash-sante-is-now-available-on-ai-gateway-for-free) - Vercel Changelog (Zachary Chen, Jerilyn Zheng), September 4, 2026.
5. [Ling 3.0 Flash Sante API, Pricing and Playground](https://vercel.com/ai-gateway/models/ling-3.0-flash-sante) - Vercel AI Gateway, accessed September 5, 2026.
6. [Hugging Face Hub API, models by author inclusionAI matching "Ling-3.0"](https://huggingface.co/api/models?author=inclusionAI&search=Ling-3.0&limit=50) - Hugging Face, accessed September 5, 2026 (no Sante repository; direct lookup of inclusionAI/Ling-3.0-flash-Sante returns an error).
7. [ModelScope model search for "Ling-3.0"](https://www.modelscope.cn/models?name=Ling-3.0&page=1) - ModelScope, accessed September 5, 2026 (the model API for inclusionAI/Ling-3.0-flash-Sante returns "record not found").
8. [Changelog](https://developer.ant-ling.com/en/docs/getting-started/changelog/) - Ant Ling developer documentation (Ant Intelligence (Hang Zhou) Technology Co., Ltd.), last updated September 4, 2026.
9. [Pricing](https://developer.ant-ling.com/en/docs/models/price/) - Ant Ling developer documentation, last updated September 4, 2026.
10. [Models: Ling](https://developer.ant-ling.com/en/docs/models/ling/) - Ant Ling developer documentation, last updated September 4, 2026.
11. [Ling-3.0-Flash-Sante on Novita AI: Medical-Tuned MoE API and Pricing](https://blogs.novita.ai/ling-3-0-flash-sante-on-novita-ai-launch-pricing) - Novita AI blog (Novita AI), September 4, 2026.
12. [Ling 3.0 Flash Sante API and Playground](https://novita.ai/models/model-detail/inclusionai-ling-3.0-flash-sante) - Novita AI, accessed September 5, 2026.
13. [inclusionAI/Ling-3.0-flash model card](https://huggingface.co/inclusionAI/Ling-3.0-flash) - Hugging Face (inclusionAI), accessed September 5, 2026.
14. [Ant Group Unveils Ling-3.0-Flash Delivering Top-Tier Performance at a Fraction of the Parameter Scale](https://www.businesswire.com/news/home/20260726584441/en/) - Business Wire (Ant Group), July 27, 2026.
15. [MedXpertQA: Benchmarking Expert-Level Medical Reasoning and Understanding](https://arxiv.org/abs/2501.18362) - arXiv (Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, Bowen Zhou), January 30, 2025 (v3 June 6, 2025).
16. [HealthBench: Evaluating Large Language Models Towards Improved Human Health](https://arxiv.org/abs/2505.08775) - arXiv (Rahul K. Arora, Jason Wei, Rebecca Soskin Hicks, Preston Bowman, Joaquin Quinonero-Candela, Foivos Tsimpourlas, Michael Sharman, Meghan Shah, Andrea Vallone, Alex Beutel, Johannes Heidecke, Karan Singhal), May 13, 2025.
17. [HealthBench Professional: Evaluating Large Language Models on Real Clinician Chats](https://arxiv.org/abs/2604.27470) - arXiv (Rebecca Soskin Hicks, Mikhail Trofimov, Dominick Lim, Rahul K. Arora, Foivos Tsimpourlas, et al.), April 30, 2026.
18. [DiagnosisArena: Benchmarking Diagnostic Reasoning for Large Language Models](https://arxiv.org/abs/2505.14107) - arXiv (Yakun Zhu, Zhongzhen Huang, Linjie Mu, Yutong Huang, Wei Nie, Jiaji Liu, Shaoting Zhang, Pengfei Liu, Xiaofan Zhang), May 20, 2025 (v5 August 17, 2026).
19. [Assessing and Enhancing Large Language Models in Rare Disease Question-answering](https://arxiv.org/abs/2408.08422) - arXiv (Guanchu Wang, Junhao Ran, Ruixiang Tang, Chia-Yuan Chang, Yu-Neng Chuang, Zirui Liu, Vladimir Braverman, Zhandong Liu, Xia Hu), August 15, 2024.
20. [What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams](https://arxiv.org/abs/2009.13081) - arXiv (Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, Peter Szolovits), September 28, 2020.
21. [MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering](https://arxiv.org/abs/2203.14371) - arXiv (Ankit Pal, Logesh Kumar Umapathi, Malaikannan Sankarasubbu), March 27, 2022.
22. [X-LANCE/MedEthicAlign: MedES, A Human-Centric Pipeline for Aligning Large Language Models with Chinese Medical Ethics](https://github.com/X-LANCE/MedEthicAlign) - GitHub (X-LANCE), repository created May 16, 2025, accessed September 5, 2026; paper: [arXiv 2601.07954](https://arxiv.org/abs/2601.07954) (Haoan Jin, Han Ying, Jiacheng Ji, Hanhui Xu, Mengyue Wu), January 12, 2026.
23. [BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents](https://arxiv.org/abs/2504.12516) - arXiv (Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, Amelia Glaese), April 16, 2025.
24. [DeepSearchQA: Bridging the Comprehensiveness Gap for Deep Research Agents](https://arxiv.org/abs/2601.20975) - arXiv (Nikita Gupta, Riju Chatterjee, Lukas Haas, Connie Tao, Andrew Wang, Chang Liu, Hidekazu Oiwa, Elena Gribovskaya, Jan Ackermann, John Blitzer, Sasha Goldshtein, Dipanjan Das), January 28, 2026.
25. [Humanity's Last Exam](https://arxiv.org/abs/2501.14249) - arXiv (Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, et al.), January 24, 2025 (v11 July 28, 2026).
26. [Ant Ling (@AntLingAGI) post introducing Ling-3.0-flash-Fin](https://x.com/AntLingAGI/status/2093022087069958492) - X, August 27, 2026.
27. [Ant Group launches finance-tuned Ling model, plans to open-source it next week](https://technode.com/2026/08/28/ant-group-launches-finance-tuned-ling-model-plans-to-open-source-it-next-week/) - TechNode (TechNode Feed), August 28, 2026.
28. [inclusionAI/Ling-3.0-flash-Fin model card](https://huggingface.co/inclusionAI/Ling-3.0-flash-Fin) - Hugging Face (inclusionAI), repository created September 3, 2026, accessed September 5, 2026.
29. [inclusionAI: Ling 3.0 Flash Fin](https://openrouter.ai/inclusionai/ling-3.0-flash-fin) - OpenRouter, listing created August 27, 2026, accessed September 5, 2026.
30. [Ant Ling (@AntLingAGI) post announcing Ling-3.0-flash-VL](https://x.com/AntLingAGI/status/2095935971556782372) - X, September 4, 2026, 18:04 UTC.
31. [Bing News search for "Ling-3.0-flash-Sante"](https://www.bing.com/news/search?q=%22Ling-3.0-flash-Sante%22&format=rss) - Bing News, accessed September 5, 2026 (no results); [Hacker News search for "Ling-3.0-flash-Sante"](https://hn.algolia.com/api/v1/search?query=Ling-3.0-flash-Sante&tags=story) - Hacker News (Algolia), accessed September 5, 2026 (no results).