Ant Lab AI Models: The Ling Family Behind Ant Group (2026 Guide)

What Is Ant Lab?

Ant Lab is the name the community uses for InclusionAI — Ant Group's dedicated AI research lab. Ant Group is one of China's largest fintech companies and the parent of Alipay, which makes the Ling models a rare case of a payments giant competing at the open-model frontier.

The lab launched in February 2025, publicly inspired by DeepSeek's example: a research lab showing that frontier-scale AI doesn't have to be a closed, capital-only game. InclusionAI's stated mission — "AGI as humanity's shared milestone, not a privileged asset" — translates into a concrete strategy: open weights, open research, and a rapid release cadence. By November 2025, the lab was being profiled as one of the fastest-rising open-model teams in the world, and it now maintains its own open-source ecosystem (model trackers, RL training suites like AReal, and the agent framework AWorld).

💡 Why it matters for this blog: Ant's bet is the same one the local-AI community made years earlier — that efficiency beats raw scale. Its models activate a tiny fraction of their parameters per token, which is the philosophy behind the MoE "A3B" models we cover in the open-source capability ranking and the sub-12B power ranking. Ant just applies it at 100B+ scale.

The Model Families: Ling, Ring, and Ming

Family	What it is	Example releases
Ling	General language models — the main series, MoE-first	Ling Plus → Ling 2.0 → Ling-2.5-1T → Ling-2.6-Flash → Ling-3.0-Flash
Ring	Reasoning models — chain-of-thought specialists	Ring 2.0, Ring-2.5-1T
Ming	Multimodal models — vision and beyond	Ming variants alongside 2.x releases

The architecture story is consistent across the family: sparse Mixture-of-Experts from day one. The first public release, Ling Plus (April 2025), was already a 293B sparse MoE, and every generation since has pushed the active-parameter ratio further — culminating in Ling-3.0-Flash activating just 5.1B of its 124B parameters per token.

The Release Timeline: 2025 → 2026

Feb 2025 — InclusionAI founded, DeepSeek-inspired, open-models mission.
Apr 2025 — Ling Plus, 293B sparse MoE — first public release.
Jun 2025 — Reasoning-model paper; lab continues building on the MoE-first approach.
Jul 2025 — Ling 1.5.
Sep 2025 — Ling & Ring 2.0 in three sizes: mini (16B total / 2B active), flash (100B total / 6B active), and the 1T primary model (1T total / 50B active), with a technical report on scaling RL to frontier models.
Feb 2026 — Ling-2.5-1T and Ring-2.5-1T, two trillion-parameter models under open licenses.
Apr 2026 — Ling-2.6-Flash (104B / 7.4B active, MIT): "state-of-the-art for its size" on agent benchmarks; a 10-point jump on the Artificial Analysis Intelligence Index over the previous Flash.
Jul 2026 — Ling-3.0-Flash (124B / 5.1B active): hybrid-reasoning MoE built for production-scale agents, free on OpenRouter through early August.

The Recent 2026 Models, Up Close

Model	Released	Total / Active	Context	License	Built for
Ling-3.0-Flash	Jul 2026	124B / 5.1B	256K	Open weights (HF)	Production-scale agents, hybrid reasoning
Ling-2.6-Flash	Apr 2026	104B / 7.4B	262K	MIT	Token-efficient agent workloads
Ling-2.6-1T	2026	1T / ~50B	262K	Open weights	Flagship quality — leads Ant's lineup on July 2026 leaderboards
Ling-2.5-1T / Ring-2.5-1T	Feb 2026	1T each	—	Open licenses	Language + reasoning at trillion-parameter scale

🌩️ Ling-3.0-Flash — The Agent Model (Jul 2026)

The current flagship of the Flash line. A hybrid-reasoning MoE — it can think fast or reason deeply within one model — with 124B total parameters and only 5.1B active per token. Native hybrid-linear attention (part MLA, part linear attention) is what makes the tiny active count possible, and it natively handles 256K context. Ant positions it squarely at production-scale agents, and it launched free on OpenRouter through early August 2026 so teams could evaluate it without risk.

⚡ Ling-2.6-Flash — The Efficiency Breakthrough (Apr 2026)

The release that put Ant on the efficiency map. 104B total / 7.4B active under an MIT license, with a hybrid linear-attention architecture (1:7 MLA + Lightning attention). Ant reported it as state-of-the-art for its size on AI-agent benchmarks, and it scored a 10-point jump on the Artificial Analysis Intelligence Index over the previous Flash. Independent tracking measured it using ~15M tokens across the full AI-Analysis test suite versus 40–100M for comparable models — the token-efficiency story in numbers.

🏔️ Ling-2.6-1T — The Trillion-Parameter Flagship

Ant's largest open model: 1T total parameters with ~50B active, 262K context, and a technical-report trail that openly documents the challenges of scaling reinforcement learning to frontier scale. It appears on the July 2026 open-source leaderboards (Onyx) as the strongest entry in Ant's lineup. Not a model you self-host — it's the reference point the Flash models are distilled toward.

🔬 Ling-2.5-1T & Ring-2.5-1T — The Open-Science Pair (Feb 2026)

Two trillion-parameter models released together under open licenses: Ling for language and Ring for reasoning. The February 2026 release signaled Ant's long-term commitment to open science at frontier scale — and set up the efficiency work that followed in the Flash line.

What Makes Ant's Models Different

Efficiency as the strategy, not a feature. Most labs ship big models and then add a small one. Ant's Flash line is engineered around active-parameter efficiency from the architecture up — hybrid linear attention instead of pure transformer attention.
Token efficiency for agents. AI agents burn tokens on tool calls, retries, and long reasoning traces. Ant's models are measured on tokens-per-task, not just accuracy — the ~15M vs 40–100M token gap on the same agent test suite is the headline stat.
Open research culture. The lab publishes technical reports on RL scaling, shares training infrastructure (AReal, AWorld), and maintains the llm-oss-landscape tracker — the Chinese-lab version of "show your work."
Fintech-grade reliability goals. Coming from a payments company, the models are aimed at workloads where correctness and cost at scale matter — agents, tool use, and business workflows rather than chat demos.

How to Run Ant Models

Hosted API: Ling-3.0-Flash on OpenRouter (free launch window through early Aug 2026) and Ant's own developer portal at developer.ant-ling.com.
Weights: the inclusionAI namespace on Hugging Face hosts the full family, including GGUF and quantized community builds.
Locally: the Flash models are efficient per token, but 104–124B total parameters still means roughly 60–80GB at Q4 — a serious workstation or a quantized build on a 48GB+ GPU. The 1T models are not self-hostable in practice. For the local-first stacks this blog covers, Ant's smaller cousins (Qwen, Gemma, Phi — see the sub-12B ranking) remain the pragmatic picks.

🚀 Context for the local-AI reader

Ant's efficiency research is worth studying even if you never run a 124B model: the same hybrid-attention and active-parameter ideas are how local MoE models like GPT-OSS and GLM Flash punch above their size on your hardware. For a full map of who's leading open-weight AI, see the capability ranking — and for the quantization math behind running any of these, our Q4_K_M vs Q8_0 guide.

Frequently Asked Questions (FAQ)

What is Ant Lab?

Ant Lab is the shorthand for Ant Group's AI research lab, InclusionAI — the team behind the Ling, Ring, and Ming model families. Ant Group is the fintech company behind Alipay, and it launched InclusionAI in February 2025 with a DeepSeek-inspired mission to build open models with AGI as a shared milestone.

What models has Ant Lab released?

Three families: Ling (general language models), Ring (reasoning models), and Ming (multimodal). The timeline runs from Ling Plus (293B sparse MoE, April 2025) through Ling 2.0 (September 2025), Ling-2.5-1T and Ring-2.5-1T (February 2026), Ling-2.6-Flash (April 2026), and the current Ling-3.0-Flash (July 2026).

What is Ling 3.0 Flash?

Ant's newest release (July 2026): a hybrid-reasoning Mixture-of-Experts model with 124B total parameters but only 5.1B active per token, built for production-scale agents. It uses native hybrid-linear attention, supports 256K context, and was offered free on OpenRouter through early August 2026.

What is the best Ant Lab model in 2026?

For agent workloads and token efficiency, Ling-3.0-Flash (124B-A5.1B) is the standout — it activates a fraction of its parameters per token, keeping cost and latency low. For maximum raw quality, the Ling-2.6-1T flagship (1T parameters, 262K context) leads Ant's lineup on the July 2026 open-source leaderboards.

Are Ant Lab models open source?

Yes — open weights are the strategy. Ling-2.6-Flash ships under MIT, the Ling-2.5-1T and Ring-2.5-1T came out under open licenses, and all are published on Hugging Face under the inclusionAI namespace. Check each model card for the exact license terms.

Can I run Ant models locally?

The Flash models' tiny active-parameter count makes them unusually efficient per token, but with 104–124B total parameters they still need serious hardware — roughly 60–80GB at Q4 quantization. The 1T flagship is not realistic to self-host. For most people, the hosted API (OpenRouter, Ant Ling API) is the practical path.