Ling-3.0-flash: What We Now Know About Ant Group's Open-Weight Fast-Tier MoE

Ant Group’s InclusionAI lab has officially released Ling-3.0-flash — announced July 24, 2026, and open-sourced under the MIT license on August 7 — and the picture has changed a lot since the preview we published here in July. It’s a native hybrid-reasoning mixture-of-experts model with 124B total parameters and just 5.1B active per token, built as the fast, cheap tier of the Ling family. The two numbers that changed everything: Artificial Analysis now scores it at 38 on the Intelligence Index (up 24 points from the previous fast-tier generation, Ling-2.6-flash) and places it on the open-weights Pareto frontier for intelligence versus total parameters. Weights are live on Hugging Face and ModelScope.

When we first covered this model on July 24, the honest summary was: API-only, no weights, no license statement, no independent benchmark. All four of those statements are now out of date. Ant open-sourced the weights under MIT on August 7, and Artificial Analysis has since published a full independent evaluation. This update revises the original brief accordingly — the “unverified” labels that dominated the July post now apply to a much narrower set of claims, chiefly Ant’s peak-speed figures, rather than the model as a whole.

TL;DR. Ling-3.0-flash is Ant Group’s fast-tier open-weight MoE: 124B total / 5.1B active, MIT license, 256K native context (262K as served). Artificial Analysis gives it an Intelligence Index of 38 — up 24 points from prior-gen Ling-2.6-flash and on the open-weights Pareto frontier for intelligence vs total parameters — with measured output around 368 tokens/sec. Weights are on Hugging Face and ModelScope under MIT. Still vendor-only: the 1,100+ tokens/sec peak-throughput claim, the sub-100ms time-to-first-token figure, and the model-card benchmark table. First-party API pricing is $0.075 in / $0.22 out per 1M tokens (80% off on cache hits); the launch free tier ended August 3.

Key takeaways

• It’s the fast/cheap tier, not the flagship. The prior generation’s 1T-parameter flagship remains InclusionAI’s largest model; Ling-3.0-flash is the smaller, faster sibling aimed at high-volume text and tool-calling.

• Weights are open under MIT now. The weights landed on Hugging Face (inclusionAI/Ling-3.0-flash) and ModelScope on August 7 under the MIT license — a reversal of the July “API-only” picture.

• It finally has an independent score. Artificial Analysis measures an Intelligence Index of 38, up 24 from Ling-2.6-flash, and puts the model on the open-weights Pareto frontier for intelligence vs total parameters.

• Speed is partly measured now. AA clocks roughly 368 tokens/sec output and a ~1.98s time-to-first-token; Ant’s headline 1,100+ tokens/sec peak figure is still vendor-only.

• Pricing is set, and the free launch is over. The first-party API lists $0.075 in / $0.22 out per 1M tokens with an 80% cache-hit discount; the launch free tier ended August 3.

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

A note on how to read this update: this is a revision of the July 24 preview, written after the weights went open and the first independent scores appeared. Every specification, benchmark, and price below is labeled by source. Where Ant Group is the only source — the peak-speed claims and the model-card benchmark table — we say so plainly. Where Artificial Analysis has measured something independently, we cite that.

What we actually know

The architecture is now fully documented on the model card. Ling-3.0-flash uses a native hybrid-linear attention stack — Kimi Delta Attention (KDA) and Gated MLA layers alternating at a 5:1 ratio (35 KDA + 7 MLA), with KDA fine-grained diagonal gating — on top of a 1/64 sparse MoE (512 routed experts, 8 activated per token). That is how it reaches 124B total parameters with only 5.1B active. The context window is 256K natively, served at 262,144 tokens in the inference recipes, and Ant claims the architecture scales to 1M. Maximum output length is not disclosed. The model is text-only with tool-calling support; multimodal input lives in the separate Ming line.

Two weight formats are published by the lab — BF16 (roughly 255GB) and FP8 (roughly 128GB) — and community quantized variants are already linked from the model card. The lab’s own deployment recipes target SGLang and vLLM.

Availability: open weights under MIT

On August 7, Ant Group’s InclusionAI team open-sourced the weights under the MIT license on Hugging Face (inclusionAI/Ling-3.0-flash) and ModelScope. That is a permissive, commercially usable license — the same one Ling 2.0 and Ling 2.6 shipped under — and it means you can self-host, fine-tune, and deploy Ling-3.0-flash yourself rather than renting it through an API. The model is also callable through Ant’s own developer API and several third-party platforms. For a fast-tier model at this price, the more interesting question is whether to self-host (FP8 runs in roughly 128GB) or stay on an API.

Independent scores: an AA Intelligence Index of 38

The single biggest change from the July post is that independent numbers now exist. Artificial Analysis has a page for Ling-3.0-flash and measures it at 38 on the Intelligence Index — 24 points above the previous fast-tier generation, Ling-2.6-flash (14), and level with MiMo-V2.5 and Qwen 3.6 27B while activating a fraction of their parameters. Artificial Analysis also places the model on the open-weights Pareto frontier for intelligence versus total parameters: no open-weights model with fewer total parameters scores higher. The flash-tier open-weights leader on AA, DeepSeek V4 Flash, still scores higher (Index 52 at max reasoning effort).

The agentic and long-context results are directionally strong as well. On AA’s τ3-Bench Banking suite, Ling-3.0-flash scores 27%, second among flash-tier open-weights models behind DeepSeek V4 Flash (39%). On GDPval-AA v2 it reaches an Elo of 1108, up from 545 for Ling-2.6-flash. Ant trained the model across 10,000+ interactive environments (vendor-reported) and pitches it as an execution node for agent workflows — fast, cheap, and disposable, with the heavy reasoning left to a larger flagship.

One caveat on sourcing: the benchmark table on Ant’s model card — SWE-Bench Pro 56.6, SWE-bench Multilingual 72.4, MathArena AIME 2026 93.2, HLE 22.7 — is vendor-reported and has not been independently reproduced yet. Treat it as the lab’s own claims, in the same category as the peak-speed figures below.

Speed: measured, then claimed

The speed story is now two different stories. Independently, Artificial Analysis measures roughly 368 tokens/sec output and a ~1.98s time-to-first-token on the first-party API — genuinely fast for a 5.1B-active MoE, and enough to rank the model near the top of its class for speed. Separately, Ant Group claims an average output rate above 1,100 tokens/sec on specified GPU configurations and sub-100ms time-to-first-token with a cluster-level hierarchical caching layer built on SGLang HiCache and Mooncake. That second set of numbers is vendor-only — measured on hardware and batch setups Ant controls, with no independent re-run published. For planning, use the AA figure; treat 1,100+ tok/s as a marketing ceiling to verify on your own workload.

Pricing: cheap, and the promo is over

Artificial Analysis lists the first-party API at $0.075 per 1M input tokens and $0.22 per 1M output, with cache hits at $0.015 (−80%) — all vendor-set pricing. The launch free tier (500k free tokens/day, free API access) ended August 3, and Ant ran a temporary 75%-off period on its Ling Studio through August 31, 2026, after which list pricing applies. Even at full list, $0.075/$0.22 puts Ling-3.0-flash firmly in the cheap tier for a model with an Index of 38.

At prices this low, the switching cost matters more than the per-token price. OrcaRouter exists to make that switch cheap: one API for 200+ models, provider list prices passed through at 0% markup, and automatic failover across providers — so when a newly-open model like this looks good on paper, you can test it against your real workload without a second contract, and fall back to a different model behind the same key if it under-delivers on a specific task.

The Ling / Ring / Ming family

Ling-3.0-flash does not exist in isolation — InclusionAI organizes its releases into three named families, each aimed at a different job. Ling is the general-purpose MoE line for everyday text generation and tool-calling, and Ling-3.0-flash sits at its fast/cheap end, a step below the previous generation’s 1T-parameter flagship. Ring is the reasoning-focused line, built for multi-step chain-of-thought. Ming is the multimodal line. If your task needs heavier reasoning or image 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.

Who it’s for: three scenarios

1. High-volume, latency-sensitive text or tool-calling pipelines

This is the model’s home turf. With an independent Index of 38, an MIT license, and roughly 368 tok/s measured, the fast-tier bet is now testable rather than hypothetical — you can run your own evaluation set through it, or pull the weights and self-host. Just don’t build around the vendor’s 1,100+ tok/s ceiling until you’ve measured it on your workload.

2. Teams that want long context on a budget

A 256K native context (262K served) with a stated path to 1M, at $0.075/$0.22, is an attractive combination for retrieval-heavy or document-processing work. The model card’s long-context numbers are vendor-reported, so shortlist it against established long-context options and verify on your own corpus before committing.

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

The July question — would InclusionAI stay open? — now has an answer: yes, MIT, and fast. Ling-3.0-flash landing on the AA Pareto frontier at 5.1B active is a data point for anyone tracking how Chinese labs are competing on efficiency rather than raw parameter count.

What’s still unverified

A short list, now that the big gaps are closed. The vendor’s peak-speed claims (1,100+ tok/s, sub-100ms TTFT) have no independent re-measurement. The model-card benchmark table is vendor-reported. The FP4/INT4 variants and the single-DGX-Spark deployment path are vendor-stated. And on knowledge, AA’s Omniscience Index shows the improvement over the prior generation came largely through abstention — hallucination fell (97% → 44%) while accuracy rose only slightly (16% → 18%) — so don’t mistake the headline Index jump for an across-the-board knowledge leap.

When not to use it

Skip Ling-3.0-flash for multimodal work — that is the Ming line. Skip it if you need a heavy reasoning flagship rather than a fast executor — that is Ring’s lane. If you plan to self-host, budget for the real hardware: BF16 is roughly 255GB, FP8 roughly 128GB. And if a claim matters to you, verify it — the peak-speed and long-context numbers are Ant’s until an independent lab re-runs them.

FAQ

Is Ling-3.0-flash open weight?

Yes. The weights were open-sourced on August 7 under the MIT license, on Hugging Face (inclusionAI/Ling-3.0-flash) and ModelScope. That is a permissive, commercially usable license — the same one Ling 2.0 and Ling 2.6 shipped under.

How does Ling-3.0-flash perform on benchmarks?

Artificial Analysis measures an Intelligence Index of 38, up 24 points from Ling-2.6-flash, and places the model on the open-weights Pareto frontier for intelligence vs total parameters. The benchmark table on Ant’s model card (for example SWE-Bench Pro 56.6) is vendor-reported and has not been independently reproduced.

How fast is it really?

Artificial Analysis measures roughly 368 tokens/sec output with a ~1.98s time-to-first-token on the first-party API. Ant claims 1,100+ tokens/sec peak throughput and sub-100ms TTFT on specified GPU configs — those are vendor figures with no independent re-measurement.

What does Ling-3.0-flash cost?

AA lists the first-party API at $0.075 per 1M input tokens and $0.22 per 1M output, with an 80% cache-hit discount. The launch free tier ended August 3, and a temporary 75%-off period on Ling Studio runs through August 31, 2026.

How big is the context window?

256K natively, served at 262,144 tokens; Ant claims the architecture scales to 1M. Maximum output length is not 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 are separate families for different jobs, not tiers of one model.

Is Ling-3.0-flash the same as the previous 1T flagship?

No. Ling-3.0-flash is the newer, smaller fast-tier release (124B total / 5.1B active). The prior generation’s 1T-parameter flagship remains the bigger model.

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

More defensible than it was in July, now that there is an independent score and an MIT license. Run your own evaluation set first, verify the vendor speed claims against your workload, and decide between the API and self-hosting (FP8 runs in roughly 128GB). The remaining vendor-only numbers are narrow enough to plan around.

Bottom line

Ling-3.0-flash graduated from “spec sheet and vendor claims” to “open weight and independently scored” in two weeks. An AA Intelligence Index of 38 at 5.1B active parameters — on the open-weights Pareto frontier, MIT-licensed, at $0.075/$0.22 — makes it one of the most efficient open-weight models you can point at right now, and one you can actually run yourself. The caveats are narrower than they were: the peak-speed and model-card figures are still Ant’s until an independent lab reproduces them. For high-volume text and tool-calling at this price, it has earned a real evaluation.