InclusionAI, the open-model research arm connected to Ant Group, released Ling 3.1 Flash on September 29 — a model with 560 billion total parameters that currently costs nothing to use at all. That combination is unusual enough to be worth a close look: most models this large launch with meaningful per-token pricing from day one, not a free trial window.
There's a catch worth understanding before you build anything production-serious on top of it, though: "free" here has a published expiration date, and the open-weight release InclusionAI has pointed toward hasn't happened yet.
Quick summary: InclusionAI released Ling 3.1 Flash on September 29, 2026 — a hybrid-reasoning, sparse Mixture-of-Experts model with a reported 560 billion total parameters and 25 billion active per token. It ships with a 262,144-token context window at launch, built for coding and tool-using agents, with InclusionAI indicating it plans to eventually unlock a full one-million-token window. The model is free to use through a two-week trial ending October 13, 2026, available via Vercel AI Gateway, OpenCode, and Kilo Code. InclusionAI has indicated it plans to publish the model's weights as open source once that trial concludes — but as of launch, the weights are not yet public, and post-trial pricing has not been announced.
The Numbers
Ling 3.1 Flash is a sparse Mixture-of-Experts model with a reported 560 billion total parameters and 25 billion active per token — meaning only a fraction of that total gets used on any given request, which is what keeps inference cost and latency manageable despite the headline size. It's described as a hybrid reasoning model, aimed specifically at coding and tool-using agent work rather than general conversation. At launch, the context window is capped at 262,144 tokens (roughly 256K), though the underlying architecture is reportedly designed for up to a million-token window once InclusionAI unlocks it fully (Source: OpenRouter).
Free — For Now
The free access runs through a two-week trial window, available via Vercel AI Gateway, OpenCode, and Kilo Code, and is scheduled to end October 13, 2026. InclusionAI has indicated that once the trial concludes, it plans to both unlock the full million-token context window and publish the model's weights as open source. Neither of those has happened yet as of this writing, and no post-trial API pricing has been announced — so the honest status right now is a free trial with stated future intentions, not a confirmed open-source release (Source: AIToolsReview).
Where It Sits Next to Other Recent Open Models
LumiChats covered Xiaomi's MiMo-V2.6-Flash in late September — a similarly-positioned sparse MoE model at 309 billion total parameters and 15 billion active, priced at $0.14 input / $0.28 output per million tokens once its own free-access windows end, and already MIT-licensed and open-weight on Hugging Face at launch. Ling 3.1 Flash is larger on paper (560B vs. 309B total parameters) but, unlike MiMo-V2.6-Flash, isn't open-weight yet — it's reachable only through the hosted trial right now. Parameter count alone isn't a reliable proxy for real-world quality, and neither model's benchmark scores here have been independently verified against the other, so this is a comparison of what's confirmed about access and scale, not a capability ranking.
| Model | Total / Active Params | Context Window | Status at Launch |
|---|---|---|---|
| Ling 3.1 Flash | 560B / 25B | 262,144 tok (target up to 1M) | Free trial through Oct 13, 2026; weights not yet open |
| MiMo-V2.6-Flash | 309B / 15B | 1,048,576 tok | MIT-licensed, open-weight on Hugging Face at launch |
If you're planning to build anything on Ling 3.1 Flash, check current pricing and license status directly before you commit production traffic to it — the free window has a stated end date of October 13, 2026, and what replaces it hasn't been announced.
Who Should Actually Try This
- Developers building coding or tool-using agents who want to test a 560B-parameter model's agentic performance before any per-token cost applies.
- Teams already evaluating Xiaomi's MiMo-V2.6-Flash or similar open/free-tier models who want another data point before committing to an inference provider.
- Anyone curious about frontier-adjacent Chinese open-model releases, given InclusionAI's stated intent to open-source the weights once the trial period ends.
01Is Ling 3.1 Flash open source?
Not yet. It's currently accessible only through a free hosted trial via Vercel AI Gateway, OpenCode, and Kilo Code. InclusionAI has indicated it plans to publish the weights as open source once that trial ends, but that hasn't happened as of this writing.
02When does the free access end?
InclusionAI's stated trial window runs two weeks from launch, ending October 13, 2026. What pricing or access model follows hasn't been announced.
03How big is Ling 3.1 Flash compared to other open models?
It's reported at 560 billion total parameters with 25 billion active per token, larger on paper than Xiaomi's MiMo-V2.6-Flash (309B total / 15B active), though parameter count alone doesn't reliably predict real-world output quality.
04What is Ling 3.1 Flash actually built for?
InclusionAI positions it for coding and tool-using agent workloads specifically, with a 262,144-token context window at launch and a stated goal of unlocking up to a million tokens later — rather than as a general consumer chat model.
If you want to see how a model like Ling 3.1 Flash actually performs against established options on your own prompts, rather than taking a launch-week spec sheet at face value, LumiChats gives you one subscription with 40+ AI models in one place — so you can judge a new release for yourself before deciding whether it's worth building around.
