Yi-Large
Last reviewed
May 16, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 3,823 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 3,823 words
Add missing citations, update stale details, or suggest a clearer explanation.
Yi-Large is a closed-source large language model developed by Chinese artificial intelligence company 01.AI (零一万物, Língyi Wànwù), founded by Kai-Fu Lee. Released on May 13, 2024, Yi-Large was the company's first proprietary flagship model, positioned at the top of the Yi family alongside the open-source Yi-1.5 series unveiled on the same day. It was offered exclusively through 01.AI's developer platform rather than as a downloadable artifact, and on debut it placed in the top tier of public chatbot leaderboards including the LMSYS Chatbot Arena, where 01.AI publicized a 7th-place finish that put the model in close competition with frontier systems from OpenAI, Anthropic and Google DeepMind.
Yi-Large arrived during what the Chinese technology press called the "price war" of mid-2024, when 01.AI, Alibaba Cloud, Baidu, ByteDance and DeepSeek cut API rates by an order of magnitude within a few weeks of each other. Yi-Large was priced at 20 RMB per million tokens (roughly $2.80 at the time), a fraction of what equivalent models from OpenAI cost in the same window. Inside 01.AI's lineup, Yi-Large served as the high-capability ceiling for paying API customers, while smaller siblings Yi-Large-Turbo, Yi-Medium, Yi-Vision and Yi-Spark covered budget, multimodal and lightweight workloads. The product line was reshaped six months later by the launch of Yi-Lightning, a Mixture-of-Experts successor optimized for speed and cost, and then by 01.AI's broader 2025 strategic pivot away from large-scale pre-training.
This article focuses on Yi-Large itself. Sister articles cover the parent company in 01.AI, the open-source models in Yi, and the speed-optimized successor in Yi-Lightning.
01.AI was formally incorporated in May 2023 by Kai-Fu Lee, a Taiwanese-born computer scientist who earned his doctorate in speech recognition at Carnegie Mellon University and went on to lead Microsoft Research Asia, Google China and the venture capital firm Sinovation Ventures. Lee began assembling the founding team in March 2023, describing large language models as a "historical opportunity" for China's technology industry. The company reached unicorn status in November 2023 after a funding round backed by Alibaba Cloud, Tencent, Xiaomi and Sinovation Ventures valued the eight-month-old startup at more than $1 billion.
From the outset, 01.AI pursued a two-track strategy: open-source releases under the Apache 2.0 license to seed adoption among researchers and downstream developers, and closed proprietary models behind an API for paying enterprise customers. The open track was anchored by Yi-34B and Yi-6B, released in November 2023, which briefly held the top spot among pre-trained base models on the Hugging Face Open LLM Leaderboard. The closed track did not become public until Yi-Large in May 2024.
01.AI held a launch event in Beijing on May 13, 2024 to announce both the Yi-1.5 open-source upgrade and the proprietary Yi-Large API. Kai-Fu Lee framed the announcement as 01.AI "officially entering" the global frontier of closed models, pointing to AlpacaEval 2.0 and SuperCLUE benchmark results in which Yi-Large was claimed to outperform GPT-4 on Chinese-language tasks and trail only GPT-4 Turbo on the English Length-Controlled Win Rate metric. The same event introduced the rest of the closed API family (Yi-Large-Turbo, Yi-Medium, Yi-Medium-200K, Yi-Vision and Yi-Spark) and disclosed that an even larger Mixture-of-Experts model code-named Yi-XLarge was already in training.
Yi-Large sits within a broader family of models built and trained by 01.AI from scratch. The table below summarises the major releases through to the closed-source proprietary tier introduced in May 2024.
| Model | Release date | Size | Type | License | Notes |
|---|---|---|---|---|---|
| Yi-6B | November 2023 | 6B | Base / Chat | Apache 2.0 | Bilingual base; 200K extended context variant available |
| Yi-34B | November 2023 | 34B | Base / Chat | Apache 2.0 | Topped Hugging Face Open LLM Leaderboard on release |
| Yi-VL-6B / Yi-VL-34B | January 2024 | 6B / 34B | Vision-language | Apache 2.0 | LLaVA-style multimodal extension |
| Yi-9B | March 2024 | 9B | Base / Chat | Apache 2.0 | Depth-upscaled from Yi-6B |
| Yi-1.5-6B / 9B / 34B | May 13, 2024 | 6B, 9B, 34B | Base / Chat | Apache 2.0 | Continued pre-training on 500B additional tokens |
| Yi-Large | May 13, 2024 | Not officially disclosed | Closed-source flagship | Proprietary, API only | LMSYS Arena top tier on launch |
| Yi-Large-Turbo | May 13, 2024 | Not disclosed | Closed-source API | Proprietary | Speed and price-optimized variant of Yi-Large |
| Yi-Medium / Yi-Medium-200K | May 13, 2024 | Not disclosed | Closed-source API | Proprietary | Mid-tier; long-context variant supports 200K tokens |
| Yi-Vision | May 13, 2024 | Not disclosed | Closed-source multimodal API | Proprietary | Image-text understanding for API customers |
| Yi-Spark | May 13, 2024 | Not disclosed | Closed-source API | Proprietary | Low-cost lightweight tier |
| Yi-Coder-1.5B / 9B | September 2024 | 1.5B, 9B | Open-source code | Apache 2.0 | 128K context, 52 programming languages |
| Yi-Lightning | October 2024 | Not disclosed (MoE) | Closed-source API | Proprietary | Mixture-of-Experts, hybrid attention |
The "Yi" name itself comes from 一, the Chinese character for the number one, an extension of the company's 零一万物 ("zero-one, all things") naming. The open-source side of the family was documented in detail in the technical report "Yi: Open Foundation Models by 01.AI" (arXiv:2403.04652, March 2024), which described the architecture, data curation pipeline and benchmark methodology used across Yi-6B, Yi-34B and their derivatives. Yi-Large is not the subject of that paper; the closed model's internals were summarised in product blog posts, press briefings and a small set of disclosed benchmark numbers rather than a full technical report.
01.AI did not publish a complete architecture paper for Yi-Large in the way it later did for Yi-Lightning. Public statements have, however, confirmed several high-level design points and ruled out others.
Kai-Fu Lee and 01.AI representatives described Yi-Large at launch as a "super-large" dense transformer language model trained at "tens of billions" of parameters, without committing to a specific count in any official release. Press coverage in May 2024 commonly cited estimates in the 70 to 100 billion parameter range, but those figures originated with analysts and were never confirmed by the company. Outside observers should therefore treat any single number as an estimate. What 01.AI did disclose publicly was that Yi-Large used the same overall architectural recipe as the open-source Yi-34B base model, which means a decoder-only transformer with grouped-query attention, SwiGLU feed-forward activations and rotary position embeddings as detailed in the open-source Yi technical report. The open-source Yi-34B configuration uses 60 layers with a hidden size of 7,168, 56 query attention heads and 8 key-value heads.
Yi-Large was distinguished from the open-source family by larger parameter scale, a more extensive bilingual training corpus and a separate alignment pipeline tuned for API customers, rather than by a fundamentally new architecture. 01.AI also confirmed that Yi-Large was a dense model and not a Mixture-of-Experts system; the MoE design was reserved first for the unfinished Yi-XLarge and later for the productized Yi-Lightning released in October 2024.
Context window length for Yi-Large was not disclosed in standardized terms. The closely related Yi-Medium-200K variant introduced on the same day extended context to 200,000 tokens for long-document workloads, while the long-context Yi-1.5 open-source models shipped with 4K, 16K and 32K variants. Yi-Large itself was offered at a single context length through the API for general-purpose dialog and reasoning rather than as a long-document tool.
Details of Yi-Large's training run were never published in a peer-reviewed paper, and 01.AI has been more reserved about closed-model internals than open-source releases. What is on the record falls into a few categories.
First, the model was trained from scratch on 01.AI's own GPU cluster, following the same general data pipeline philosophy as the open-source Yi models. The Yi technical report describes a heavily filtered bilingual corpus of approximately 3.1 trillion tokens with cascaded deduplication and quality filtering, and 01.AI told reporters at launch that Yi-Large was trained on a substantially expanded version of that pipeline rather than on third-party data. The exact token count for Yi-Large was not disclosed.
Second, the model's alignment used the same recipe family employed for Yi-34B-Chat. The open-source paper describes a deliberately small (less than 10,000 example) supervised fine-tuning corpus emphasizing quality over quantity, followed by direct preference optimization. 01.AI representatives have indicated that Yi-Large used a similar two-stage approach with a larger and higher-quality preference dataset, but the company has not released sample counts, annotation procedures or reward model details. There is no public confirmation of reinforcement learning from human feedback specifically; the open-source Yi report uses DPO instead, and Yi-Large is widely believed to follow suit.
Third, Yi-Large was reported by Chinese trade press to have been trained on the company's bare-metal cluster of NVIDIA H800 GPUs (the export-control-compliant version of the H100 sold in China), although 01.AI did not publish the exact size of the cluster or the FLOPs budget consumed by the run. Kai-Fu Lee later cited the training cost of frontier models in his explanation for 01.AI's 2025 strategic pivot, saying that "only tech giants can bear the costs of training super-large models." Public statements at the time of Yi-Large's launch put the company's total compute spend at approximately $100 million, including data, salaries and electricity for the cluster.
Because Yi-Large was never released as weights, its public performance record is built from a combination of leaderboard rankings (LMSYS Chatbot Arena, AlpacaEval 2.0, SuperCLUE) and benchmark numbers reported by 01.AI during the May 2024 launch event. The headline result was an LMSYS Arena ranking that placed the model in the top ten globally on debut and at or near the top among Chinese-developed models.
The figures below are drawn from 01.AI's launch materials, the SuperCLUE leaderboard archive and reporting by KrASIA, Pandaily and SCMP. Numbers from external sources are noted where they differ from the company's own claims.
| Benchmark | Yi-Large | Notes |
|---|---|---|
| LMSYS Chatbot Arena (overall, May 2024) | Top 10 globally; 7th place at debut | Ranking was for an early API checkpoint; later revisions placed it slightly lower as new entries appeared |
| AlpacaEval 2.0 Win Rate | Highest overall Win Rate of any model evaluated at the time | 01.AI claim, May 13, 2024 launch slides |
| AlpacaEval 2.0 Length-Controlled Win Rate | 2nd, behind GPT-4 Turbo | 01.AI claim, May 13, 2024 launch slides |
| SuperCLUE-Math6 | 1st among Chinese-language LLMs | 01.AI claim, May 2024 |
| SuperCLUE Chinese Hard (combined) | Outperformed GPT-4 across six SuperCLUE evaluation datasets, per 01.AI | KrASIA, May 2024 |
| MMLU (5-shot) | Comparable to GPT-4 (mid 80s range, per company materials) | Exact figure not independently verified |
| AlignBench | Top among Chinese-developed models on Chinese alignment categories | SuperCLUE / launch materials |
| HumanEval | Reported as comparable to GPT-4-class models | No standardized figure released |
A few caveats are worth flagging. The Arena ranking was for an early Yi-Large API checkpoint and the leaderboard position fluctuated as new models entered. The AlpacaEval 2.0 numbers were generated by an LLM-as-judge methodology, which is sensitive to length and style; the LC (Length-Controlled) variant corrects for some of that bias and ranked Yi-Large second rather than first. Some of the launch claims (notably the GPT-4 head-to-head on SuperCLUE) compared Yi-Large to specific GPT-4 checkpoints rather than to the latest GPT-4 Turbo or GPT-4o models then in circulation, a fact that several outlets noted in their coverage. 01.AI has not, to date, released raw chat logs or per-task breakdowns sufficient for full third-party reproduction.
Where later reporting placed Yi-Large in the broader leaderboard, it was generally described as a top-tier but not single-model leader. In a August 2024 LMSYS update, Yi-Large was ranked in the upper third of all evaluated systems and within the top three Chinese-developed models alongside Qwen 2 and DeepSeek V2.
Yi-Large was distributed through 01.AI's developer platform, hosted at platform.lingyiwanwu.com (also accessible at platform.01.ai). Customers signed up for an API key, charged a wallet with RMB credit and called the model through an OpenAI-compatible chat-completions endpoint.
The platform exposed several closed-source models in parallel, each with its own price tier and target workload. The table below summarises the lineup that launched alongside Yi-Large on May 13, 2024.
| API model | Target workload | Approximate launch price (per 1M tokens) |
|---|---|---|
| Yi-Large | High-capability reasoning, complex tasks | 20 RMB (about $2.80) |
| Yi-Large-Turbo | Faster, lower cost variant of Yi-Large | 12 RMB |
| Yi-Medium | General-purpose mid-tier | 2.5 RMB |
| Yi-Medium-200K | Long-document reasoning (200K tokens) | 12 RMB |
| Yi-Vision | Multimodal image and text | 6 RMB |
| Yi-Spark | Lightweight low-cost workloads | 1 RMB |
The 20 RMB per million token price for Yi-Large was less than one third of the contemporary GPT-4 Turbo rate, which 01.AI repeatedly used as a comparison point in its launch communications. The model was billed per input and output token separately, with an additional context-cache discount for repeat prefixes that the platform documentation introduced later in 2024. Free-tier credits were initially offered to new developer accounts as an onboarding promotion.
A few weeks after launch, 01.AI joined what the Chinese tech press called the "big model price war" of mid-2024, in which Alibaba Cloud, Baidu, ByteDance and DeepSeek cut their API prices by 90% or more within days of each other. 01.AI matched competitor cuts for Yi-Large-Turbo and Yi-Medium but kept Yi-Large itself at its launch price, framing the flagship as a premium tier rather than a commodity workload model.
Yi-Large was also distributed through 01.AI's consumer-facing chatbot product Wanzhi (万知), which positioned itself as a productivity assistant for Chinese users, and through partnerships with several Chinese cloud providers for enterprise deployments.
The table below contrasts Yi-Large with three contemporary flagship models that were widely tested against it in mid-2024. Where a figure is not publicly disclosed for Yi-Large, the entry reads "not disclosed" rather than reusing a competitor's number.
| Property | Yi-Large (May 2024) | GPT-4 Turbo (April 2024) | Claude 3 Opus (March 2024) | Qwen 2-72B-Instruct (June 2024) |
|---|---|---|---|---|
| Developer | 01.AI | OpenAI | Anthropic | Alibaba Cloud |
| Availability | Closed, API only | Closed, API only | Closed, API only | Open weights (Apache 2.0) |
| Architecture | Dense decoder-only transformer | Closed; reported to be MoE | Closed; not disclosed | Dense decoder-only transformer |
| Parameter count | Not officially disclosed (estimates 70B-100B) | Not disclosed | Not disclosed | 72B (open weights) |
| Context window | Not standardized in launch materials | 128K | 200K | 32K (Instruct), 128K via YaRN |
| Languages | Strong Chinese and English bilingual | Strong multilingual, English-led | Strong multilingual, English-led | Strong Chinese and English bilingual |
| LMSYS Arena rank at launch | Top 10 globally, top among Chinese models | Top 3 globally | Top 3 globally | Top 10 globally for open models |
| API price (per 1M tokens, USD-equiv at launch) | About $2.80 | About $10 input / $30 output | About $15 input / $75 output | Free if self-hosted; cents through Alibaba Cloud |
| Notable strength | Chinese-language tasks, SuperCLUE benchmarks | Tool use, vision, ecosystem maturity | Long-form writing, long context, refusal calibration | Strong open-weights flagship; bilingual |
This comparison is best read as a snapshot of May to June 2024. By the second half of 2024, DeepSeek's V2 and V3 releases, Baichuan's Baichuan 4 and Qwen 2.5 had reshuffled the relative positions of Chinese frontier models, and by 2025 the picture had shifted again with DeepSeek-R1, Qwen 3 and 01.AI's own Yi-Lightning.
On pricing in particular, Yi-Large entered an environment where Chinese closed APIs were already dramatically cheaper than their Western counterparts, and the gap widened over the rest of 2024 as Chinese cloud providers competed on per-token cost. The headline 01.AI claim that Yi-Large undercut GPT-4 Turbo by roughly two thirds was accurate at launch, but the competitive frame shifted within weeks rather than months as 01.AI's domestic peers cut their own prices.
Yi-Large was 01.AI's high-water mark for flagship pre-training. Within about six months of its launch the company had begun to restructure around a different strategic bet, and within a year it had publicly stepped back from frontier pre-training altogether.
In October 2024, 01.AI released Yi-Lightning, a Mixture-of-Experts model that prioritized speed and cost over raw scale. Yi-Lightning placed sixth on the LMSYS Chatbot Arena leaderboard on debut and used a hybrid attention design that the company said reduced key-value cache memory by up to 82.8%. In commercial terms, Yi-Lightning replaced Yi-Large as the model 01.AI actively promoted for enterprise workloads, although Yi-Large remained available through the API.
In mid-December 2024, reports emerged in the Chinese technology press that 01.AI had reassigned its pre-training algorithm and infrastructure teams. Late-December reporting indicated that members of the pre-training group had received offers from Alibaba's Tongyi (通义) division, while infrastructure engineers had received offers from Alibaba Cloud. The reporting was widely interpreted as a sign that 01.AI was effectively exiting frontier pre-training. Kai-Fu Lee initially denied that 01.AI had sold its team or assets, calling some of the reports "vicious slander," but acknowledged a significant strategic shift.
In January 2025, Alibaba Cloud and 01.AI announced a joint laboratory focused on industrial AI models. In a public letter, Lee explained the move by writing that "only tech giants can bear the costs of training super-large models" and that "spending on more GPUs to train large models isn't the way forward for survival" for a startup of 01.AI's size. He framed large models as "teacher models" whose capabilities would be distilled into smaller, more deployable application models for 01.AI's enterprise customers.
By March 2025, 01.AI had formally stopped pre-training large language models and pivoted to selling tailored AI business solutions, in many cases on top of DeepSeek's open-source models rather than 01.AI's own. The company reported revenue exceeding 100 million yuan (approximately $13.7 million) for 2024, with roughly 70% from enterprise clients in gaming, finance and energy and the remainder from international consumer products. Several senior staff departed during the transition, including a vice president who moved to ByteDance and the lead engineer for the company's multimodal effort.
As a practical matter, this means Yi-Large is the last fully self-trained 01.AI dense flagship of significant scale. Yi-Lightning followed a few months later as a Mixture-of-Experts successor, and no comparably large dense model was released by 01.AI in 2025 or, as of mid-2026, in 2026. Yi-Large continued to be served through the platform.lingyiwanwu.com API into 2025, but enterprise sales attention shifted to Yi-Lightning and to bespoke application stacks built around external open-source models.
Reception to Yi-Large at launch was generally positive within the Chinese technology press and cautious internationally. Pandaily, KrASIA and the South China Morning Post all carried prominent coverage of the May 13 announcement, focusing on the SuperCLUE results, the AlpacaEval 2.0 win rate and the API price. The Pandaily headline ("01.AI Releases A Trillion Parameter Closed-Source Large Model 'Yi-Large'") attached a parameter count that 01.AI itself had not confirmed, an issue that recurred in subsequent secondary sources.
Independent benchmarking on the LMSYS Chatbot Arena placed Yi-Large in the top tier of all evaluated systems on debut, with a particularly strong showing on Chinese-language prompts. Western analysts at outlets including The Information and Bloomberg noted that the Arena performance was real and impressive, but cautioned that Yi-Large was being judged against a fast-moving frontier in which OpenAI, Anthropic and Google were releasing new models almost monthly. By the end of summer 2024, several frontier models had moved past Yi-Large on the Arena leaderboard while Chinese peers including Qwen 2 and DeepSeek V2 closed in from behind.
The API price drew sustained attention. Several Chinese outlets characterized 01.AI's 20 RMB per million token rate as the moment that turned the 2024 "price war" into open competition; rivals announced cuts within days. Whether Yi-Large's pricing was sustainable was an open question at the time, and analysts including those at GGV Capital and Citic Securities expressed doubt that any Chinese AI startup could indefinitely undercut OpenAI on price while matching it on capability.
Reception in retrospect, after 01.AI's late-2024 restructuring and 2025 pivot, has been more mixed. Some commentators have read Yi-Large as a technical success that nonetheless failed to find a profitable commercial niche; others have read it as a credible flagship that was overtaken by faster-moving rivals before it could establish a defensible market position. Kai-Fu Lee's own framing, in interviews and the January 2025 letter explaining the pivot, was that Yi-Large had validated 01.AI's training capabilities while clarifying that the company's long-term value would come from applications rather than from continuing to chase frontier scale.