What is Ling-3.1-flash?
Ling-3.1-flash is a language model from inclusionAI, the AI research arm of Ant Group, released with open weights on Hugging Face in October 2026. It is a hybrid-linear mixture-of-experts model with 560 billion parameters, 25 billion active per token, and a 1-million-token context. Most of its layers use a linear attention design, with one standard attention layer in eight, which keeps long contexts cheaper to process. It reads and writes text, and reasoning can be switched on or off.
The weights are under the MIT licence. Ant recommends SGLang for serving, with a reference setup on eight H200 GPUs and published recipes for H100, H800, B200 and GB300. Alongside it, Ant introduced LingCode, a coding-agent harness built on the open-source Pi agent, which it may open-source later.
inclusionAI serves it through its own API and OpenRouter. Artificial Analysis lists inclusionAI's price at $0.30 per million input tokens and $0.90 per million output tokens, while OpenRouter lists it at $0.075 and $0.225.
How much does Ling-3.1-flash cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | MIT licence |
| inclusionAI API | $0.30 in / $0.90 out per 1M tokens | Cached input $0.06, per Artificial Analysis |
| OpenRouter | $0.075 in / $0.225 out per 1M tokens |
Open weights under MIT. inclusionAI API: $0.30 input / $0.90 output per 1M tokens, cached $0.06, per Artificial Analysis; OpenRouter lists $0.075 / $0.225.
Is Ling-3.1-flash worth it?
Ling-3.1-flash's selling point is speed. Artificial Analysis measured about 211 to 219 tokens a second, roughly four times MiMo-V2.6-Pro's 57 and well above most models in its class. It scores 41 on Artificial Analysis's Intelligence Index, double its predecessor's 20 and a few points behind the leading open models (MiMo-V2.6-Pro 46, GLM-5.3 45, Kimi K3 44).
The catch is verbosity. It used about 218 million output tokens to complete Artificial Analysis's index, more than twice the median, so its cost per index task ($1.01) is higher than its low token price suggests and well above MiMo-V2.6-Pro's $0.13. For short, latency-sensitive answers that matters less; for long agent runs it adds up.
The licence is clean MIT, and the hybrid-linear design is built for long contexts, so self-hosting on a single eight-GPU H200 node is realistic. Ant's own benchmark tables are published only as images, and it names coding, office, healthcare and finance tasks as strengths without independent numbers to back them.
The gap is transparency about hosted use. inclusionAI publishes a training-content summary for the model, but its API's privacy and retention terms are hard to find, and the two listed API prices differ by a factor of four. Self-host it, or route through a provider whose terms you know.
Where it earns its place
A product team building a responsive assistant
Answer quickly with near-frontier open-model quality
At over 200 tokens a second, a 500-token answer streams in about 2.5 seconds after a 1.8-second first chunk. Turn reasoning off for short replies to save tokens.
A company self-hosting under a permissive licence
Run a long-context model on its own GPUs
Serve it with Ant's SGLang image on eight H200 GPUs with tensor and expert parallelism, or follow the B200 and GB300 recipes. MIT puts no conditions on commercial use.
A developer comparing low-cost open models
Pick between speed and cost per task
Ling-3.1-flash is fastest but verbose; MiMo-V2.6-Flash or DeepSeek-V4.1-Flash cost less per completed task on Artificial Analysis's index.
Before you commit
- Setup
- Call it through OpenRouter or inclusionAI's API, or serve the weights with the day-0 SGLang image. Ant recommends temperature 1.0, top_p 0.95, top_k 20 and enable_thinking on.
- Integrations
- SGLang with MTP speculative decoding and HiCache; works with Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw, per Ant; OpenRouter.
- Data and privacy
- Weights under MIT, with a public training-content summary. inclusionAI's API privacy and retention terms were not found on its model or OpenRouter pages; self-host for sensitive data.
- Cost at scale
- At $0.30/$0.90, 1,000 agent tasks of 500K input and 30K output tokens cost about $177 uncached, before its above-average reasoning output. At OpenRouter's listed price, about a quarter of that.
- Limits
- 1M-token context; text only. About 1.8 seconds to first chunk and heavy reasoning output (9.5 seconds of thinking on Artificial Analysis's long-prompt test).
What's good
- Over 200 tokens a second on Artificial Analysis's tests
- MIT licence with published SGLang recipes for H100 to GB300
- Scores 41 on Artificial Analysis, double its predecessor
- Hybrid-linear attention for cheaper 1M-token contexts
What's not
- Uses over twice the median output tokens on Artificial Analysis's index
- $1.01 per index task despite low token prices
- Text only
- API data terms not easy to find; listed prices disagree
Best for
Teams that want a fast, permissively licensed open model for latency-sensitive chat and agents, self-hosted or via OpenRouter.
Ling-3.1-flash alternatives
Ling-3.1-flash: common questions
- Is Ling-3.1-flash worth it?
- Fast and MIT-licensed with near-frontier open-model scores, but it burns tokens thinking and its data terms are not published. We rate it 4.0 out of 5.
- How much does Ling-3.1-flash cost?
- Open weights under MIT. inclusionAI API: $0.30 input / $0.90 output per 1M tokens, cached $0.06, per Artificial Analysis; OpenRouter lists $0.075 / $0.225.
- What are the downsides of Ling-3.1-flash?
- Uses over twice the median output tokens on Artificial Analysis's index; $1.01 per index task despite low token prices; text only; aPI data terms not easy to find; listed prices disagree.
- Who is Ling-3.1-flash best for?
- Teams that want a fast, permissively licensed open model for latency-sensitive chat and agents, self-hosted or via OpenRouter.
- How much does Ling-3.1-flash cost for a team or in production?
- At $0.30/$0.90, 1,000 agent tasks of 500K input and 30K output tokens cost about $177 uncached, before its above-average reasoning output. At OpenRouter's listed price, about a quarter of that.
- Is Ling-3.1-flash safe for company data?
- Weights under MIT, with a public training-content summary. inclusionAI's API privacy and retention terms were not found on its model or OpenRouter pages; self-host for sensitive data.
- What are the best Ling-3.1-flash alternatives?
- The closest alternatives we have reviewed are MiMo-V2.6-Pro, Step 5 Preview, GLM-5.3-Flash and DeepSeek-V4.1-Flash.
Answers checked Oct 25, 2026.
Make Ling-3.1-flash? Show this rating on your site
