InclusionAI
Last reviewed
May 16, 2026
Sources
31 citations
Review status
Source-backed
Revision
v3 ยท 5,528 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
May 16, 2026
Sources
31 citations
Review status
Source-backed
Revision
v3 ยท 5,528 words
Add missing citations, update stale details, or suggest a clearer explanation.
| inclusionAI | |
|---|---|
![]() | |
| Type | AI research initiative / open-source AI lab |
| Industry | Artificial intelligence |
| Founded | February 2025 |
| Headquarters | Hangzhou, China |
| Key people | Zhengyu He (CTO of Ant Group), Richard Bian (product and growth lead), Wu Yi, Le Zhenzhong, Cai Wei, Shen Chunhua, Yang Ming, Zheng Da |
| Parent | Ant Group |
| Owner | Ant Group |
| Products | Ling text LLMs, Ring reasoning models, Ming multimodal models, LLaDA diffusion language models, AReaL reinforcement learning system, ASearcher, AWorld, Inclusion Arena |
| Website | inclusion-ai.org |
inclusionAI is an open-source artificial general intelligence (AGI) research initiative established by Ant Group, the financial-technology affiliate of the Alibaba ecosystem.[1] The lab develops and openly releases large language models, multimodal models, reinforcement learning systems, agent frameworks, and evaluation platforms, with model weights and training recipes distributed primarily through Hugging Face, ModelScope, and GitHub.[2] Founded in February 2025 in the aftermath of DeepSeek's breakthrough release of R1, inclusionAI has emerged as one of the most prolific Chinese open-weight model publishers, scaling from sub-20-billion-parameter checkpoints in early 2025 to a family of trillion-parameter mixture of experts (MoE) models by the end of 2025 and the spring of 2026.[3][4]
The organization positions itself as a community-driven hub for projects originating from Ant Group's research and engineering teams, with the stated mission of treating AGI as a shared global milestone rather than a proprietary asset. inclusionAI releases its flagship models under permissive licenses, most often MIT, and accompanies them with technical reports, training infrastructure code, and reproducibility notes.[5] Its model families include the Ling general-purpose LLMs, the Ring thinking-oriented reasoning models, the Ming multimodal omni-models, the LLaDA diffusion language models, and the DR-Venus edge-scale deep research agents.[2]
inclusionAI operates as Ant Group's primary public face for foundation-model research, sitting alongside but distinct from the parent company's product-facing AI groups such as the team behind the Alipay AI assistant and the AQ healthcare app. The initiative was publicly framed in interviews and corporate communications as Ant Group's most concentrated push toward general-purpose intelligence, with the lab reporting directly to the office of Ant Group's chief technology officer, Zhengyu He.[6][7]
The lab's design philosophy rests on three explicit principles articulated on its own website. The first is open by default, under which models, training code, and selected datasets are released to the public so that researchers outside Ant Group can reproduce results. The second is ecosystem oriented, with formal partnerships and frequent informal collaboration with academic groups such as the Institute for Interdisciplinary Information Sciences at Tsinghua University. The third is start small, a pragmatic stance favoring controlled experiments with small MoE proxies that fit predicted scaling laws before scaling up to multi-hundred-billion or trillion-parameter checkpoints.[5]
As of mid-2026, inclusionAI's Hugging Face organization page lists more than 130 model artifacts and 20 datasets, with roughly 60 named team members and active collections covering the Ling 2.0, Ling 2.5, Ling 2.6, Ring 2.5, Ring 2.6, Ming, and LLaDA series.[2] The lab also operates the Inclusion Arena live evaluation leaderboard and contributes to community projects such as vLLM and SGLang for efficient inference deployment.[5]
Ant Group's research arm, Ant Research, had been working on large language models and reinforcement learning systems for several years before the formal launch of the inclusionAI brand, with internal projects feeding into Alipay and the company's enterprise customers. Public references to inclusionAI as a named open-source initiative began appearing in early 2025, following Ant Group's broader pivot toward an AI-first corporate strategy.[6]
In interviews with the Interconnects newsletter, product and growth lead Richard Bian dated the initiative's formal founding to February 2025 and described the project as a direct response to the success of DeepSeek R1, which had demonstrated that a focused Chinese lab could ship a frontier-grade open-weight reasoning model with limited compute. Ant Group's senior leadership concluded that a dedicated open-source AGI program inside the company was both feasible and strategically necessary, and inclusionAI was established as the vehicle for that program.[6]
The first major release under the inclusionAI banner was the Ling Plus and Ling Lite pair, announced in March 2025. Ling Lite is a 16.8-billion-parameter MoE model with 2.75 billion activated parameters per token, while Ling Plus uses 290 billion total parameters with 28.8 billion activated parameters. Ant Group disclosed that the models were trained primarily on domestically produced accelerators from Alibaba and Huawei, with limited use of Nvidia hardware, and reported that training one trillion tokens cost approximately 6.35 million yuan, equivalent to roughly 880,000 US dollars at the time of release. The company claimed a 20 percent cost reduction relative to an Nvidia-only baseline and benchmarked the models as broadly comparable to similarly sized open Chinese models.[8][9]
In April 2025, Ant Group launched a recruitment program called Plan A, designed to attract graduate students and early-career researchers to the company's AGI mission. As part of the program, Ant Group publicly profiled twelve senior researchers who would serve as mentors. The list, reported by the South China Morning Post, included Wu Yi, a former OpenAI researcher and Berkeley PhD; Le Zhenzhong, a Carnegie Mellon graduate who had worked at Google AI; Cai Wei, a Stanford graduate and former senior Google engineer behind parts of Google Image Search and the Cloud Vision API; Shen Chunhua, a statistical machine learning researcher with an h-index above 130; Yang Ming, a founding member of Meta's FAIR research lab; and Zheng Da, who runs Ant's graph computing laboratory. Zhengyu He, profiled as a Georgia Tech PhD known for record-setting maximum-flow algorithms on GPUs, was reaffirmed as Ant Group's chief technology officer and as the executive sponsor of the AGI program.[7]
inclusionAI gave one of its first major academic appearances at the International Conference on Learning Representations (ICLR) 2025 Expo, where the team presented its open reinforcement learning training stack and early agent work. The presentation centered on AReaL, the asynchronous RL framework jointly developed with Tsinghua University's Institute for Interdisciplinary Information Sciences. A stable v1.0 of AReaL followed later in 2025, with a subsequent boba and boba-2 release that reported roughly 2.77x faster training than synchronous baselines on agentic reinforcement learning workloads.[10][11]
A second wave of releases began on 10 September 2025 with the open-sourcing of the Ling 2.0 architecture and its first members, Ling-mini-2.0 and Ling-flash-2.0. The series was the first to adopt the Ling Scaling Laws, a set of empirical relationships that the inclusionAI team uses to predict optimal MoE configurations, training schedules, and activation ratios. The 2.0 series committed to a 1/32 activation ratio (256 routed experts and one shared expert per layer) and to FP8 mixed-precision training across all sizes.[12]
On 30 September 2025, inclusionAI released Ring-1T-preview, a preview checkpoint of a trillion-parameter reasoning model, on Hugging Face. The release was widely reported as the first open-source trillion-parameter thinking model, predating the comparable open release of any Western lab.[13][14] On 9 October 2025, Ant Group announced the general release of Ling-1T, the trillion-parameter base model from which Ring-1T was derived, and the Ling-1T technical materials confirmed that pre-training had used more than 20 trillion tokens and roughly 50 billion activated parameters per token.[15][16] The full Ring-1T post-trained model followed on 14 October 2025, with the arXiv paper Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model (arXiv 2510.18855) detailing the icepop reinforcement learning algorithm and the in-house ASystem RL framework.[17]
In February 2026, Ant Group released Ling-2.5-1T and Ring-2.5-1T together. Ling-2.5-1T extended context to one million tokens and was tuned to reach frontier reasoning quality with much smaller output budgets than peer thinking models, reportedly matching frontier results on AIME 2026 using only around 5,890 output tokens versus the 15,000 to 23,000 tokens typical of competing models. Ring-2.5-1T introduced a hybrid linear-attention architecture, which Ant Group described as the world's first such configuration at trillion-parameter scale, and posted gold-medal-tier scores on the International Mathematical Olympiad 2025 (35 of 42) and the China Mathematical Olympiad 2025 (105 of 126).[18][19]
In the spring of 2026, inclusionAI rolled out the Ling 2.6 and Ring 2.6 series. Ling-2.6-1T uses a hybrid architecture combining Multi-Latent Attention (MLA) with linear attention, has approximately one trillion total parameters with roughly 63 billion activated parameters, and supports a one-million-token context window. The model focuses on what the team calls quick thinking, emphasizing short, high-quality completions for execution-style tasks rather than long chain-of-thought reasoning.[20] Ring-2.6-1T, released roughly half a day before this article was written, is the matching reasoning-oriented sibling at the trillion scale.[2]
Parallel to the Ling and Ring series, inclusionAI built out the Ming multimodal family and the LLaDA diffusion language model family. Ming-Omni, described in a paper published 11 June 2025, integrated dedicated encoders for vision, audio, and video tokens with the Ling MoE backbone and was characterized in independent coverage as the first open-source model matching GPT-4o's modality coverage. Ming-flash-omni 2.0 followed in February 2026 with 100 billion total parameters and 6 billion activated parameters, and a Ming-Omni-TTS announcement landed in March 2026.[21][22] LLaDA2.0-Uni, released in late 2025, brought together unified visual understanding, image generation, and image editing in a 16-billion-parameter discrete diffusion MoE backbone, while LLaDA 2.1 in February 2026 added accelerated text-diffusion via token editing.[23]
inclusionAI's published mission statement frames the lab's goal as providing a sustainable and responsible intelligence option for learners and builders, with a secondary slogan of bringing small and usable intelligence to the world.[5] Internally the team describes language models as infrastructure kernels analogous to operating-system primitives, on top of which downstream agent ecosystems, retrieval systems, and product surfaces can be built.
The three operating principles repeatedly cited in lab communications are open by default, ecosystem oriented, and start small. Open by default means that model weights, tokenizers, and training-time configurations are normally released alongside published checkpoints, with permissive licenses where legal and policy constraints permit. Ecosystem oriented means that the lab favors collaboration with academic groups (Tsinghua University is the most frequent partner), with inference framework communities such as vLLM and SGLang, and with downstream open-source projects rather than a closed product-stack-first posture. Start small means that scaling decisions are gated through small MoE proxy runs in what the team calls the Ling Wind Tunnel, a fixed-recipe set of experiments used to fit power-law predictions for loss, activation balance, and expert imbalance at far larger scales.[5][12]
inclusionAI's catalog is organized into several parallel families. The text-only Ling series provides general-purpose foundation models; the Ring series adds reasoning-oriented post-training to Ling bases; the Ming series adds multimodal encoders, decoders, and routers on top of Ling; LLaDA explores discrete diffusion as an alternative paradigm; and DR-Venus targets edge-scale deep research agents trained on small, focused datasets.
The Ling series is the lab's text-only foundation line, built from the ground up on sparse MoE primitives. The family scaled rapidly from sub-20-billion-parameter checkpoints in early 2025 to multiple trillion-parameter checkpoints by 2026.
| Ling model | Total parameters | Activated parameters | Context | Public release | Notes |
|---|---|---|---|---|---|
| Ling-Lite | 16.8B | 2.75B | 16K | March 2025 | Trained on domestic Chinese chips, MIT license |
| Ling-Plus | 290B | 28.8B | 64K | March 2025 | Reported 20% cost reduction vs Nvidia baseline |
| Ling-1.5 | mid-100B class | not disclosed | n/a | July 2025 | Internal milestone bridging 1.0 to 2.0 architectures |
| Ling-mini-2.0 | 16B | 1.4B (789M non-embedding) | 128K | 10 September 2025 | 300+ tokens/s on H20, matches 7-8B dense quality |
| Ling-flash-2.0 | 100B | 6.1B (4.8B non-embedding) | 128K (YaRN) | September 2025 | Matches roughly 40B dense quality |
| Ling-1T | ~1T | ~50B per token | 128K | 9 October 2025 | Largest known FP8-trained foundation model; 20T training tokens |
| Ling-2.5-1T | ~1T | ~50B per token | 1M | 15 February 2026 | Frontier reasoning quality at ~5,890 output tokens on AIME 2026 |
| Ling-2.6-1T | ~1T | ~63B per token | 1M | Spring 2026 | Hybrid MLA + linear-attention, focus on quick thinking |
| Ling-2.6-flash | 107B | not fully disclosed | extended | Spring 2026 | Flash-tier successor to Ling-flash-2.0 |
Ling 2.0 and later releases share an architectural recipe with 256 routed experts plus a single shared expert per layer, an aux-loss-free sigmoid routing balancer, QK-Norm, partial Rotary Positional Embeddings (RoPE), Multi-Token Prediction (MTP) loss, and FP8 mixed-precision training throughout. The training corpus exceeds 20 trillion tokens, beginning with a 4K context-length stage and progressively raising the share of math, code, and reasoning-dense sources to nearly half of the data mix as context windows are extended.[12][16]
The Ring series adds reasoning-oriented post-training on top of Ling bases, with extensive use of verifiable-reward reinforcement learning. Each Ring checkpoint corresponds to a Ling base of the same generation.
| Ring model | Base | Total / activated | Context | Public release | Notable benchmark results |
|---|---|---|---|---|---|
| Ring-lite | Ling-Lite | 16.8B / 2.75B | 16K | mid-2025 | Lightweight reasoning checkpoint |
| Ring-V2 | Ling 2.0 base | n/a | n/a | mid-2025 | Reasoning MoE LLM |
| Ring-1T-preview | Ling-1T base | ~1T / ~50B | 64K | 30 September 2025 | 92.6% on AIME 2025; first open trillion thinking model |
| Ring-1T | Ling-1T base | ~1T / ~50B | 64K (extendable to 128K via YaRN) | 14 October 2025 | 93.4% AIME 25, 94.69 CodeForces, 81.59 Arena-Hard v2; 4 IMO 2025 problems at silver standard |
| Ring-2.5-1T | Ling-2.5-1T | ~1T trillion-scale | 1M | 15 February 2026 | First hybrid linear-attention thinking model; 35/42 IMO 2025, 105/126 CMO 2025 |
| Ring-2.6-1T | Ling-2.6-1T | ~1T | 1M | Spring 2026 | Reasoning sibling of Ling-2.6-1T |
The Ring papers, especially Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model (arXiv 2510.18855), detail two engineering contributions widely cited in subsequent literature. The first is the icepop algorithm, a stabilization technique for RL on MoE models that mitigates the gradient explosions that can occur when expert routing shifts mid-training. The second is ASystem, an in-house reinforcement learning framework with a serverless sandbox hybrid reward system that allows the team to mix verifiable mathematical rewards, code-execution rewards, and human-preference rewards inside a single training loop.[17]
The Ming series turns Ling MoE backbones into unified perception-and-generation systems. Ming-Omni, formally introduced in arXiv 2506.09344 on 11 June 2025, used the Ling MoE backbone with newly proposed modality-specific routers, the Qwen2.5 vision tower as its image encoder, and Whisper as its audio encoder.
| Ming model | Total / activated | Capabilities | Public release |
|---|---|---|---|
| Ming-lite-omni | 22B / 3B | Image, text, audio, video understanding and generation; dialect support | June 2025 |
| Ming-lite-omni 1.5 | n/a | Enhanced multimodal capabilities | mid-2025 |
| Ming-Omni | n/a | First open-source model matching GPT-4o modality coverage in independent reviews | 2025 |
| Ming-flash-omni-Preview | n/a | Preview of flash-tier omni model | 28 October 2025 |
| Ming-flash-omni 2.0 | 100B / 6B (Ling 2.0 backbone) | Speech, audio, and music in unified pipeline; continuous autoregression with diffusion-transformer head | 11 February 2026 |
| Ming-Omni-TTS | n/a | Dedicated text-to-speech head | 4 March 2026 |
The Ming family supports speech and image generation, dialect understanding, voice cloning, context-aware multimodal dialog, text-to-speech, and image editing inside a single set of weights, an unusually wide capability surface for an open-weight model release.[21][22]
LLaDA is inclusionAI's exploratory diffusion-based language model series. LLaDA2.0-Uni, released in November 2025, scaled the LLaDA diffusion approach to a 16-billion-parameter MoE backbone capable of unified visual understanding, high-fidelity image generation, and image editing, using a SigLIP-VQ tokenizer to convert images to discrete semantic tokens and a distilled diffusion decoder enabling roughly eight-step inference. LLaDA 2.1, released in February 2026, accelerated text diffusion through token editing.[23]
The DR-Venus series consists of small, four-billion-parameter checkpoints in supervised-fine-tuning and reinforcement-learning variants, aimed at edge-scale deep research agents trained with minimal data. The models are distributed as GGUF files for local CPU and consumer GPU inference, and form part of inclusionAI's broader push to demonstrate that capable agents can be built at sizes far smaller than the flagship Ling and Ring releases.[2]
In addition to model weights, inclusionAI maintains an extensive open-source infrastructure stack covering reinforcement learning training, agent runtimes, evaluation, and benchmarking.
AReaL (Ant Reasoning RL) is an open-source asynchronous reinforcement learning training system for large reasoning and agentic models, jointly developed by the AReaL team at Ant Group and Tsinghua University's Institute for Interdisciplinary Information Sciences. The system decouples rollout generation from policy updates, enabling continuous GPU utilization across both inference and training workers and avoiding the synchronous-stall bottleneck of conventional RL pipelines. The stable v1.0 release packaged the framework with a Proxy Worker abstraction that lets agent codebases connect with a single API change, and the boba-2 update reported approximately 2.77x faster training over synchronous baselines on agentic RL workloads.[10][11]
ASearcher is an open-source framework for large-scale online RL training of search agents, designed to push so-called search intelligence to expert-level performance. The repository includes templates for building customized search agents and integrates natively with AReaL for RL fine-tuning.[24]
AWorld is a runtime system for building, evaluating, and training general multi-agent assistance. The framework supplies the orchestration scaffolding for collaborative multi-agent systems and offers an evaluation harness for measuring task completion in shared environments.[25]
Inclusion Arena is a live leaderboard and open platform for evaluating large foundation models on real-world, in-production applications, launched in August 2025. Unlike traditional lab benchmarks based on static held-out test sets, Inclusion Arena collects production traffic from participating apps and uses pairwise human preference judgments and outcome metrics to rank LLMs and multimodal LLMs in the wild. The associated paper appeared on arXiv as 2508.11452.[26]
ABench is the lab's internal benchmark suite, and dInfer is an inference framework introduced alongside Ling-1T specifically for diffusion language models. Ant Group has claimed that dInfer runs up to three times faster than vLLM and approximately ten times faster than Nvidia's Fast-dLLM framework on representative diffusion workloads, although these figures come from the company's own measurements rather than independent benchmarks.[15]
inclusionAI's published technical reports and blog posts describe a distinctive engineering stack that combines aggressive use of low-precision arithmetic, asynchronous RL, and proxy-driven scaling experiments.
The Ling Scaling Laws are a set of empirical relationships fit from small MoE runs that the lab uses to choose activation ratios, expert counts, learning rates, and tokenizer parameters before launching full-scale training. The same recipe with 256 routed experts and a single shared expert per layer is shown to remain optimal across more than four orders of magnitude in compute, from roughly 16 billion to one trillion total parameters, supporting a 1/32 activation ratio across the entire family. The team refers to this experimentation regime as the Ling Wind Tunnel.[12]
Ling-1T is described by Ant Group as the largest-scale known foundation model trained entirely with FP8 mixed-precision arithmetic. A systematic FP8-versus-BF16 comparison run for roughly one trillion tokens reported a loss deviation of approximately 0.1 percent, validating end-to-end stability and yielding what the team reports as a 15-percent-plus speedup and significant VRAM savings relative to BF16 training. FP8-specific kernels and operator fusion strategies, especially across the FC1, gated function, and FC2 path of the MoE feed-forward layer, are used to keep the schedule compute-bound rather than memory-bound on heterogeneous Chinese accelerators.[16][6]
For the Ring series, inclusionAI uses verifiable-reward reinforcement learning (RLVR) in combination with reinforcement learning from human feedback (RLHF) for taste and instruction-following. The Ring-1T paper introduces the icepop algorithm to stabilize trillion-scale MoE RL training, addressing situations in which expert-routing drift would otherwise produce gradient explosions or representation collapse. The training loop runs inside ASystem, an in-house framework with a serverless sandbox hybrid reward server that lets the team mix automatic math and code rewards with model-judged preference rewards inside a single rollout. Language-level Policy Optimization (LPO), in which sentences rather than tokens or full sequences are used as the RL unit, is reported in the Interconnects interview as a stabilizing technique that improves training stability and generalization on long-form reasoning tasks.[17][6]
The early Ling Plus and Ling Lite releases were explicitly positioned as evidence that Ant Group could train competitive MoE models on a mix of domestic accelerators from Alibaba and Huawei, with limited reliance on the highest-end Nvidia chips that were subject to United States export controls. Ant Group reported training-cost figures of approximately 880,000 US dollars per one trillion tokens at that scale, representing a roughly 20 percent reduction relative to an Nvidia-only configuration. Subsequent releases including the trillion-parameter Ling-1T have continued the use of heterogeneous accelerator pools, supported by the lab's Elastic Distributed Training (EDiT) method.[8][9]
inclusionAI is led at the executive level by Zhengyu He, the chief technology officer of Ant Group. Zhengyu He holds a PhD in computer science from Georgia Tech and was profiled by South China Morning Post in 2025 for his earlier academic work on the then-fastest maximum-flow algorithm on a GPU.[7] Day-to-day, the inclusionAI organization is run by a set of product, research, and engineering leads. Richard Bian, previously the lead of the AntOSS open-source project at Ant Group, runs product and growth for inclusionAI and Ant Ling, the brand under which several model releases appear. Research leads cited in interviews include Ziqi Liu, an eight-year Ant veteran with a PhD from Shanghai Jiao Tong University, and Chen Liang, an algorithm engineer who has driven the lab's FP8 training stabilization work.[6]
The broader research bench, profiled as part of the Plan A recruitment program, includes Wu Yi (former OpenAI, UC Berkeley PhD), Le Zhenzhong (former Google AI, Carnegie Mellon), Cai Wei (former senior algorithm engineer at Google, Stanford), Shen Chunhua (statistical machine learning with an h-index above 130), Yang Ming (founding member of Meta FAIR, Northwestern PhD), and Zheng Da (Johns Hopkins PhD, head of the graph computing laboratory at Ant Technology Research Institute).[7] The lab reports roughly sixty named team members on its Hugging Face organization page as of mid-2026.[2]
Ant Group is the largest fintech company in China and the operator of Alipay, which serves roughly 800 million users in mainland China and a growing list of international payment partners. Ant Group originated as a payment business within Alibaba before being spun off as a separate company in the 2010s, and while Ant Group is not a subsidiary of Alibaba, the two companies share founding shareholders and a deeply intertwined ecosystem of cloud, e-commerce, and financial services.[27]
inclusionAI's role within this ecosystem is twofold. The lab supplies foundation models that are fine-tuned and deployed by Ant Group's product-facing AI teams, including the Alipay AI assistant, the AI Pay agentic checkout flow, the AQ AI healthcare app (which reportedly surpassed 100 million users within one year of launch and over 140 million by late 2025), and the AI Doctor Assistant built with Haodf.com.[28][29] At the same time, the lab functions as a recruiting and reputational engine for Ant Group's AGI ambitions, providing a publicly visible body of work that anchors talent acquisition, academic collaboration, and external developer adoption.
At the annual INCLUSION conference on the Bund in Shanghai, Ant Group has used the inclusionAI brand to anchor several major announcements, including the September 2025 release of Ling 2.0 and the open-sourcing of additional infrastructure components. In November 2025, Ant Group also unveiled LingGuang, a multimodal AI assistant with code-driven outputs that draws on the Ling backbone for its underlying reasoning.[30]
The inclusionAI catalog is distributed primarily through three channels. The first is Hugging Face, where the inclusionAI organization page hosts model weights, FP8 and int4 quantized variants, datasets, and community spaces. The second is ModelScope, the Chinese model repository operated by Alibaba Cloud, which mirrors most major releases for users behind the Great Firewall. The third is GitHub, where the inclusionAI organization hosts training code, inference scripts, framework repositories such as AReaL, ASearcher, and AWorld, and technical reports.[2][5]
Flagship checkpoints including Ling-1T, Ring-1T, and Ring-1T-FP8 are released under the MIT license, an unusually permissive choice for a major Chinese foundation-model release. Some smaller specialized checkpoints, including certain DR-Venus and Ming variants, are distributed under model-specific community licenses with use-case restrictions. The training-stack repositories such as AReaL are released under permissive open-source licenses; specific terms vary by repository.
inclusionAI checkpoints have appeared frequently at or near the top of open-source leaderboards for mathematics, code, and general reasoning since late 2025. Selected reported results include:
| Model | Benchmark | Score | Reference |
|---|---|---|---|
| Ring-1T-preview | AIME 2025 | 92.6% | [13] |
| Ring-1T | AIME 25 | 93.4% | [17] |
| Ring-1T | CodeForces | 94.69 | [17] |
| Ring-1T | Arena-Hard v2.0 | 81.59% (first among open models) | [17] |
| Ring-1T | IMO 2025 | 4 problems solved at silver-medal level | [17] |
| Ring-1T | ICPC World Finals 2025 | 5 problems solved | [17] |
| Ring-2.5-1T | IMO 2025 | 35/42 (gold-medal standard) | [19] |
| Ring-2.5-1T | CMO 2025 | 105/126 (above national-team cutoff) | [19] |
| Ling-1T | AIME 2025 | 70.42% with ~4,000 output tokens per problem | [15] |
| Ling-1T | 31-benchmark suite | Top in 22 of 31 non-reasoning benchmarks | [16] |
| Ling-2.5-1T | AIME 2026 | Matches frontier reasoning models at ~5,890 output tokens | [18] |
Independent coverage by VentureBeat, the DeepLearning.AI Batch, and FinTech Weekly has generally described the Ring and Ling 2.x releases as state-of-the-art among open-weight models in their respective generations, with particular emphasis on the efficient-reasoning angle (Ling 2.5 and Ling 2.6) and the trillion-parameter-MoE engineering achievement.[14][15][18]
In the Chinese open-weight ecosystem, inclusionAI sits alongside several peer labs that became prominent in the wake of DeepSeek R1. The most directly comparable peers are the Qwen team at Alibaba Cloud, the GLM team at Z.ai (Zhipu), Moonshot AI's Kimi line, DeepSeek, Tencent's Hunyuan team, ByteDance's Doubao group, and Baidu's open-source releases. By number of model artifacts published on Hugging Face, inclusionAI is one of the most prolific Chinese open-weight publishers as of mid-2026.[2][31]
Industry analysts often characterize the lab as differentiated along three axes. First, inclusionAI was among the earliest open labs to ship trillion-parameter checkpoints (Ring-1T-preview, September 2025) at a time when most Western labs kept comparable scales closed. Second, the lab leans heavily on engineering disclosure, regularly publishing detailed training reports and infrastructure code rather than weights alone. Third, the affiliation with Ant Group, and through Ant's relationships with Alipay and Alibaba, gives the lab access to large-scale deployment surfaces in finance and healthcare that most pure-research labs lack, which informs the reasoning-and-execution emphasis of Ling 2.5 and Ling 2.6.[6][20]
Competitive coverage by Stanford's HAI policy team and similar groups places inclusionAI inside what is now commonly described as the second tier of Chinese open-weight publishers, behind Alibaba's Qwen line and DeepSeek in cumulative download counts but ahead of most newer entrants in the breadth of modalities released. Chinese open-source LLM share of global usage grew from roughly 1.2 percent in late 2024 to nearly 30 percent during 2025, with Qwen and DeepSeek leading and inclusionAI's Ling and Ring series accounting for a small but rapidly growing share.[31]
Notable third-party coverage of inclusionAI includes a long-form interview in the Interconnects newsletter by Nathan Lambert (June 2025) that detailed the lab's history, philosophy, and engineering stack;[6] coverage in VentureBeat of the Ring-1T release focused on the engineering of trillion-scale MoE reinforcement learning;[17] a feature in DeepLearning.AI's The Batch on the Ling-1T release;[15] and ongoing coverage in TechNode, Pandaily, FinTech Weekly, the South China Morning Post, and Reuters of the broader Ant Group AI strategy and Plan A recruitment program.[13][7]
Independent assessments of inclusionAI's releases have raised several open questions. Benchmark figures reported by the lab, especially for in-house tools such as dInfer, have so far been published primarily by Ant Group and have not yet been broadly reproduced by external researchers.[15] The model cards for Ring-1T explicitly note known limitations including occasional identity-recognition bias, language-mixing artifacts in long generations, repetitive output in some prompts, and inference inefficiencies stemming from the use of grouped-query attention (GQA) at very long context lengths.[17] Coverage in Western tech media has additionally raised broader questions about Chinese open-weight releases, including whether they should be considered open source in the Free Software Foundation sense given that training data is generally not released alongside weights, and whether export-control-driven reliance on domestic accelerators creates lock-in risks for downstream users.
As a research initiative tied to Ant Group, inclusionAI is also subject to Chinese AI regulations, including registration requirements for generative AI services under the Cyberspace Administration of China's Provisions on the Management of Generative AI Services. inclusionAI's public communications generally frame compliance with these regulations as a baseline for safe deployment rather than as a constraint on open-source release.
Multimodal learning
Open-source artificial intelligence