Ling 3.1 Flash is a hybrid reasoning mixture-of-experts model from inclusionAI, the model team at Ant Group. inclusionAI positions it for coding, tool use, planning and long-context agent workflows — work that runs in several steps rather than a single answer.
It carries roughly 560 billion total parameters with about 25 billion active per token. That ratio is the point of the design: the model has the capacity of a very large network while only paying to run a small slice of it per token, which is what lets it answer at the latency its name suggests.
Reasoning is part of the model rather than a separate mode. According to inclusionAI, thinking is enabled by default and can be switched off, so the same model serves both deliberate multi-step work and quick extraction where the extra tokens are not worth paying for.
The model takes text and returns text, and supports tool calling, so it can be pointed at pipelines that hand it a toolset and let it work through a task. inclusionAI targets a context window of up to one million tokens; the window served here is 256K, and that is the figure listed below.
Weights are not published yet — inclusionAI has said it plans to open-source the model, but as of this writing there is no downloadable checkpoint and no evaluation from outside Ant Group. Every benchmark figure on this page is the vendor's own.
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
inclusionAI positions Ling 3.1 Flash for low-latency assistance, extraction and routine automation — work where a request arrives often and the answer has to come back quickly, and where the cost of a frontier-sized dense model would not be justified.
The published figures lean agentic rather than conversational: the model is measured on skill execution, automation, terminal work and codebase question answering, alongside domain suites in security, finance and health. Tool calling and hybrid reasoning are what make that range possible, so pipelines that hand the model a toolset and let it work through a task are the intended shape.
import requests r = requests.post( "https://api.aimlapi.com/v1/chat/completions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "inclusionai/ling-3.1-flash", "messages": [ { "role": "user", "content": "Hello!" } ] }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "inclusionai/ling-3.1-flash", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"inclusionai/ling-3.1-flash","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Output | |
No per-token charge is listed for this model in the AI/ML API model feed at the moment.
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| skillsBench | 68.7% | Breadth of discrete professional skills carried out end to end | Source | October 2, 2026 |
| AutomationBench | 52.5% | Completion of routine automation workflows | Source | October 2, 2026 |
| CyberGym | 87.9% | Security task completion in sandboxed environments | Source | October 2, 2026 |
| Finance Agent v2 | 57.9% | Financial research and analysis run as an agent | Source | October 2, 2026 |
| DRACO | 85.5% | Data research and analysis with complex operations | Source | October 2, 2026 |
| Terminal-Bench 4 | 40.4% | Autonomous shell and terminal task completion, fourth release | Source | October 2, 2026 |
| SWE-Atlas Codebase QnA | 55.9% | Answering questions about an unfamiliar codebase | Source | October 2, 2026 |
| HealthBench Professional | 65.3% | Clinical and professional health question answering | Source | October 2, 2026 |
| Intelligence | 41.1 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | October 3, 2026 |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Ling 3.1 Flash This page | Coding, tool use, planning and long-context agent workflows | |||
| Fast, low-cost reasoning and tool use at scale | ||||
| Low-cost image and video understanding |
Yes. The model was announced by inclusionAI on 30 September 2026 and is available through AI/ML API's chat completions endpoint as inclusionai/ling-3.1-flash. The model feed currently lists no per-token charge for it.
It is a mixture-of-experts model with roughly 560 billion total parameters, of which about 25 billion are active for any given token. That ratio is the point of the design: the model has the capacity of a very large network while only paying to run a small slice of it per token, which is what keeps latency low.
256K tokens as served at launch, with a maximum output of 32,768 tokens. inclusionAI has said it targets a context window of up to one million tokens; until that is actually served, 256K is the figure to plan against.
inclusionAI positions it for low-latency assistance, extraction and routine automation. Its published figures are mostly agentic rather than conversational — skill execution, automation, terminal work and codebase question answering, plus domain suites in security, finance and health — so it is aimed at pipelines that hand the model a toolset and let it work through a task.
No. Every figure on this page was reported by inclusionAI at launch and none has been reproduced by a third party. The model's weights are not published either, so independent evaluation is not yet possible. Treat the numbers as the vendor's own claims.
Ling 3.0 Flash is a far smaller mixture-of-experts model — roughly 124 billion total parameters with about 5 billion active — and it is available today. Ling 3.1 Flash raises both figures by roughly four to five times, to about 560 billion total and 25 billion active, and adds hybrid reasoning. Both are served with a 256K context window.
