Ling-3.0-flash: What We Actually Know About Ant Group's New Fast-Tier MoE (And What We Don't)

Ant Group's InclusionAI lab released Ling-3.0-flash on or around July 23, 2026 — which means it is days old as of this brief. It's a mixture-of-experts (MoE) language model with 124B total parameters and roughly 5.1B active per token (about a 24:1 sparsity ratio), built for text generation and tool-calling. It's positioned as the fast, cheap tier of the Ling family; the flagship spot still belongs to the much larger Ling-2.6-1T (1T total / 63B active). Access today is API-only — through Ant's own developer portal, through OpenRouter as inclusionai/ling-3.0-flash, and through ZenMux.

That's the entire confirmed picture, and it's worth being upfront about why this brief reads more cautious than a typical model writeup. There are no Hugging Face weights, no GitHub repository for Ling-V3, and no clear license statement — a real change from Ling 2.0 and Ling 2.6, both of which shipped as open weights under MIT. There is also no independent benchmark data of any kind: Artificial Analysis, the most commonly cited third-party evaluator for this class of model, doesn't yet have a page for it. Every performance and speed number in circulation right now traces back to Ant Group itself.

TL;DR. Ling-3.0-flash is a 124B/5.1B-active MoE model from Ant Group's InclusionAI, released within the last few days, API-only with no Hugging Face weights and an unconfirmed license. Vendor claims — 1000 tokens/sec peak throughput, sub-100ms time-to-first-token — have no independent verification yet, and there's no Artificial Analysis score for this specific model. Prior-gen Ling-2.6-flash scored an Artificial Analysis Intelligence Index of just 14 (Ling-2.6-1T scored 26), which is useful context but not a stand-in for 3.0-flash's real capability. Pricing (¥0.40 in / ¥1.20 out per 1M tokens, currently free during launch week with 500k free tokens/day) is explicitly promotional and likely to move. Treat everything here as a preview, not a verdict.

Key takeaways

• It's the fast/cheap tier, not the flagship. Ling-2.6-1T remains InclusionAI's largest model; Ling-3.0-flash is a smaller, cheaper, faster sibling aimed at high-volume use.

• No independent benchmark exists yet. Artificial Analysis has no page for Ling-3.0-flash. The only public numbers are Ant Group's own, and Ant's docs currently show no benchmark table for this model at all.

• Speed claims are vendor-only. 1000 tokens/sec peak and sub-100ms time-to-first-token come from Ant Group, with no third-party throughput test to confirm either figure.

• No open weights, license unconfirmed. There's no Hugging Face repo and no GitHub Ling-V3 project. Prior Ling generations were MIT-licensed; whether 3.0-flash will follow is unknown.

• Pricing is a launch-week promotion. ¥0.40/¥1.20 per 1M tokens on Ant's platform is currently waived, with 500k free tokens/day and a free OpenRouter listing — none of which should be assumed to hold once the promotion ends.

• It's one piece of a three-model family. InclusionAI splits its lineup into Ling (general-purpose MoE, this model), Ring (reasoning-focused), and Ming (multimodal).

A note on how to read this brief: every specification, benchmark, and price below is labeled by source. Where Ant Group is the only source, we say so plainly and hedge accordingly. Where prior-generation Ling models have independent scores, we cite them for context — not as a substitute for data on Ling-3.0-flash itself, which doesn't yet exist. This will likely need a follow-up once independent testing catches up.

What we actually know

Architecturally, Ling-3.0-flash is a sparse MoE model: 124B parameters total, with roughly 5.1B active on any given token, which is what makes it cheap and fast to run relative to its total size. The context window is 256K tokens by Ant's own documentation, with a vendor claim that it's extendable to 1M; OpenRouter's listing shows 262,144 tokens, which lines up with the 256K figure. Maximum output length isn't disclosed anywhere we could find. The model supports standard text generation and tool-calling, but no multimodal input or output has been mentioned — that capability sits with the separate Ming line.

On availability, the picture is API-only and consistent across the three routes we found: Ant Group's own developer platform, OpenRouter (listed as inclusionai/ling-3.0-flash), and ZenMux. None of these expose downloadable weights. There's no GitHub organization page for a "Ling-V3" project, and Ant's documentation doesn't state a license for this specific model — a meaningful gap, since Ling 2.0 and Ling 2.6 were both released as open weights under MIT. Until InclusionAI clarifies this, don't assume Ling-3.0-flash will end up open, and don't assume it won't.

The gaps: what's unverified

The biggest gap is benchmarks. As of this writing, Artificial Analysis — the independent evaluator most commonly cited for exactly this class of model — has no page for Ling-3.0-flash; the URL pattern that would normally host it returns a 404. That means there is currently no third-party reasoning, coding, or math score for this model, from Artificial Analysis or any other independent evaluator we could locate. Ant's own documentation compounds this: it doesn't publish a benchmark table for 3.0-flash at all, vendor or otherwise, which is unusual for a model release and worth flagging on its own.

For rough context, prior-generation Ling models do have independent scores. Ling-2.6-flash posted an Artificial Analysis Intelligence Index of 14, and the larger Ling-2.6-1T scored 26 — both well behind the leading open-weight models tracked by Artificial Analysis at the time. Ant's own vendor figures for Ling-2.6-flash claimed 61.2% on SWE-bench Verified and 73.85% on AIME2026. None of these numbers should be read onto Ling-3.0-flash directly; a new generation with a new architecture can move in either direction, and until an independent evaluator actually runs it, any capability claim for 3.0-flash is a guess dressed up as a fact.

Vendor speed claims: fast on paper, unverified in practice

Ant Group's headline performance pitch for Ling-3.0-flash isn't reasoning quality — it's raw speed. The vendor claims a peak throughput of 1000 tokens per second and a time-to-first-token under 100 milliseconds, both of which would be genuinely fast if confirmed. But "if confirmed" is doing real work in that sentence: these numbers come exclusively from Ant Group's own materials, measured under conditions Ant controls and hasn't fully disclosed (hardware, batch size, prompt length, and load all move throughput numbers substantially). No independent lab has published a re-measurement. Until someone outside Ant runs a comparable test, treat 1000 tok/s and sub-100ms TTFT as marketing figures worth checking against your own workload, not settled facts.

Pricing, and why it's a moving target

On Ant Group's own platform, list pricing for Ling-3.0-flash is ¥0.40 per 1M input tokens and ¥1.20 per 1M output tokens — cheap by design, consistent with its positioning as the fast/cheap tier. But that list price isn't actually what anyone is paying right now: Ant is running a free launch-week promotion, and separately offering 500k free tokens per day on its platform. OpenRouter's listing shows a ling-3.0-flash:free variant at $0/$0. None of this should be mistaken for a stable price. Launch promotions expire, free tiers get capped, and the ¥0.40/¥1.20 figures themselves could shift once real usage data comes in. If you're planning production spend around this model, budget against the list price, not the promotional one, and re-check before committing.

The Ling / Ring / Ming family

Ling-3.0-flash doesn't exist in isolation — InclusionAI organizes its releases into three named families, each aimed at a different job. Ling is the general-purpose MoE line, covering everyday text generation and tool-calling; Ling-3.0-flash sits at the fast/cheap end of it, with Ling-2.6-1T still the larger flagship. Ring is InclusionAI's reasoning-focused line, built for tasks that lean on multi-step chain-of-thought rather than raw throughput. Ming is the multimodal line, handling image and other non-text modalities that Ling itself doesn't touch. If your task needs heavier reasoning or multimodal input, the right InclusionAI model is probably not Ling-3.0-flash — it's built for volume and speed within the text-and-tools lane, not for stretching into the other two families' territory.

Who it's for: three scenarios

1. High-volume, latency-sensitive text or tool-calling pipelines — as a pilot, not a commitment

If your workload is dominated by throughput-sensitive text generation or tool-calling — chat, lightweight agents, high-frequency API calls — Ling-3.0-flash's positioning (small active parameter count, claimed sub-100ms TTFT, near-free launch pricing) makes it worth a pilot. The catch is that "worth a pilot" is different from "worth switching production traffic to." With zero independent benchmark data, you're the benchmark right now: run your own evaluation set through it before trusting it with anything customer-facing, and don't build cost models around the current promotional pricing.

2. Teams that need long context on a budget

A 256K context window (with a vendor claim of extendability to 1M) at a genuinely low list price is an attractive combination for retrieval-heavy or document-processing use cases, provided the model's actual reasoning quality holds up over that context length — which, again, no independent source has tested. This is a reasonable candidate to shortlist alongside established cheap long-context options, but it should be evaluated side-by-side with them rather than adopted on the strength of Ant's specs sheet alone.

3. Watchers tracking China's MoE landscape and the InclusionAI roadmap

For teams that track the competitive landscape of Chinese open and semi-open model labs rather than deploy any single model in production, Ling-3.0-flash is a useful data point: it signals InclusionAI iterating quickly on its fast tier, potentially stepping back from open weights (at least for now), and continuing to bet on a three-family structure (Ling/Ring/Ming) rather than one general model. Worth bookmarking and revisiting once benchmark and license clarity arrive.

When not to use it

Skip Ling-3.0-flash for anything where you need to cite a verified capability claim — there is no independent benchmark to point to, so any statement about its reasoning, coding, or math ability is currently unfalsifiable. Skip it if you need open weights for self-hosting, fine-tuning, or compliance reasons; there are none available, and the license for this specific model hasn't even been stated, let alone confirmed as permissive. Skip it for multimodal tasks — that's the Ming line's job, not Ling's. And be cautious about building a cost model around today's pricing: the free launch week, the 500k free daily tokens, and the free OpenRouter tier are all promotional, and the ¥0.40/¥1.20 list price it's likely to revert to hasn't itself been tested against real-world usage patterns.

FAQ

Is Ling-3.0-flash open weight?

Not currently. There's no Hugging Face repository and no GitHub Ling-V3 project as of this writing. Prior Ling generations (2.0 and 2.6) were released under MIT as open weights, but nothing confirms that Ling-3.0-flash will follow that pattern — or that it won't.

How does Ling-3.0-flash perform on benchmarks?

There's no independent benchmark for it yet — Artificial Analysis has no page for this model. Ant Group's own documentation doesn't publish a benchmark table for it either. For rough historical context, prior-gen Ling-2.6-flash scored an Artificial Analysis Intelligence Index of 14, and Ling-2.6-1T scored 26, but neither number should be assumed to carry over to this new generation.

How fast is it really?

Ant Group claims 1000 tokens/sec peak throughput and sub-100ms time-to-first-token. Both are vendor-only figures with no independent re-measurement published so far. Treat them as claims to verify on your own workload, not confirmed performance.

What does Ling-3.0-flash cost?

List pricing on Ant's platform is ¥0.40 per 1M input tokens and ¥1.20 per 1M output tokens, but it's currently free during a launch-week promotion, with 500k free tokens/day also available. OpenRouter lists a free variant as well. Expect these promotional terms to change.

How big is the context window?

256K tokens according to Ant's documentation, with a vendor claim that it's extendable to 1M. OpenRouter's listing shows 262,144 tokens, consistent with the 256K figure. Maximum output length isn't disclosed.

What's the difference between Ling, Ring, and Ming?

Ling is InclusionAI's general-purpose text and tool-calling MoE line, and Ling-3.0-flash sits at its fast/cheap end. Ring is the reasoning-focused line. Ming handles multimodal tasks. They're separate model families built for different jobs, not tiers of the same model.

Is Ling-3.0-flash the same as Ling-2.6-1T?

No. Ling-2.6-1T is InclusionAI's larger flagship (1T total / 63B active parameters) and remains a separate, bigger model. Ling-3.0-flash (124B total / ~5.1B active) is the newer, smaller, faster-tier release.

Should I use Ling-3.0-flash in production today?

Only after running your own evaluation. With no independent benchmark, an unconfirmed license, and pricing that's currently promotional, it's reasonable to pilot but premature to commit production-critical or compliance-sensitive workloads to it without your own testing and a plan for what happens when the promotional terms end.

Bottom line

Ling-3.0-flash is a genuinely new release worth tracking — a 124B/5.1B-active MoE model aimed squarely at fast, cheap, high-volume text and tool-calling work, from a lab that has shipped credible models before. But right now it's a specs sheet and a set of vendor claims, not a verified product: no independent benchmark, no open weights, no confirmed license, and pricing that's explicitly promotional. That's not a reason to dismiss it — it's a reason to pilot it cautiously, verify the claims that matter to you directly, and revisit this brief once Artificial Analysis or another independent evaluator publishes real numbers.