On September 29, Ant Group's inclusionAI lab shipped Ling-3.1-flash, successor to Ling-3.0-flash. The efficiency-focused MoE model scales to roughly 560B total parameters with about 25B activated per token — versus the previous generation's 124B total / 5.1B active, total parameters have grown more than fourfold and active parameters nearly fivefold. The official promise that came with it: open weights after the trial period.
From 124B to 560B: even the efficiency line is stacking parameters
Ling-3.0-flash made its name on "small activation for production agents": 124B total parameters with only 5.1B active kept deployment costs low. Ling-3.1-flash blows up both numbers — 560B total, 25B active — keeping the sparsity ratio around 1/22 but landing in a completely different cost class. The industry signal is blunt: pure small-activation designs have hit a quality ceiling on agent tasks. Ant positions the model for agent tasks, search, office software, and specialist applications, with an explicit thinking-mode switch aimed at long-horizon work.
On context, the stated target is up to 1M tokens, but the current two-week free trial via Vercel's AI Gateway (hosted by Novita) opens only 256K, with output capped at 32,768 tokens (TechNode: https://technode.com/2026/09/30/ant-group-launches-ling-3-1-flash-with-560-billion-parameters/ ).
A try-first, open-later release rhythm
The most interesting thing about this launch isn't the parameters — it's the sequence. The model arrived as a two-week free trial on Vercel's AI Gateway; weights, license, and official pricing were all absent at announcement, with inclusionAI only committing to open-source after the trial ends. The independent tracker LLM Releases explicitly marks "weights not yet posted" on the model page ( https://www.llm-releases.com/models/ling-3-1-flash ), and OrcaRouter's commentary is blunter: this is an "announced model" — declared, but not yet touchable.
Compare that with Ling-3.0-flash, which went straight to Hugging Face under MIT. Gathering real-world load and feedback before releasing weights looks more like a commercial-model playbook. You can read it as quality polishing — or as a probe of open-source strategy.
Read the benchmarks with a discount
Officially reported scores: an 81.0 aggregate across five agent benchmarks, Terminal-Bench 4 at 40.4%, SWE-Atlas 55.9%, HealthBench Professional 65.3%; Ant Ling's account on X adds GDPVal-AA v2.1 at 1,673 Elo and FrontierSWE at 75.16 (official announcement: https://x.com/AntLingAGI/status/2105335205741596911 ). For reference, Claude Sonnet 5.5 scores 70.6% on the same Terminal-Bench 4.0 — Ling-3.1-flash's 40.4% sits well behind the frontier tier, while HealthBench Professional 65.3% maps neatly onto Ant's own healthcare and finance verticals. Every number so far has been run by Ant alone, with no independent replication.
So what
Whether Ling-3.1-flash is worth watching comes down to what lands after two weeks: real weights, and under what license. A 560B open model puts self-hosting into multi-GPU-cluster territory, so "open source" for most developers degrades from "I can run it" to "I can fine-tune or distill it." If it truly ships under MIT, the community gets a large base model for free; if it arrives with commercial strings, this trial was pure marketing. The answer arrives within two weeks.