Yi-Large

Yi-Large is a closed-weights proprietary large language model developed by 01.AI and released in June 2024, built on a dense decoder-only transformer architecture with a 32,768-token context window and strong multilingual capability across English, Chinese, Spanish, Japanese, German, and French. It is available through 01.AI's API platform at platform.lingyiwanwu.com and through third-party inference providers including NVIDIA NIM and OpenRouter. Yi-Large held the role of 01.AI's closed-source flagship from its June 2024 release until Yi-Lightning superseded it in October 2024; as of April 2026, Yi-Large remains available via API while 01.AI's own model development has substantially wound down following the lab's December 2024 restructuring.

At a glance

• Lab: 01.AI
• Released: June 25, 2024
• Modality: Text
• Open weights: No. Yi-Large is a proprietary closed model. Several earlier Yi family members (Yi-34B, Yi-6B, Yi-9B, Yi-1.5 series) are open-weights under the Yi license. Yi-Large is not.
• Context window: 32,768 tokens
• Pricing: $3.00 per million input tokens and $3.00 per million output tokens via third-party providers at the time of release. 01.AI's own API platform (platform.lingyiwanwu.com) also provides access; rates vary by provider.
• Distribution channels: 01.AI API at platform.lingyiwanwu.com, NVIDIA NIM at build.nvidia.com, OpenRouter, Artificial Analysis

Origins

01.AI was founded in March 2023 by Kai-Fu Lee, the former president of Google China and founding director of Microsoft Research Asia. Lee framed the company's founding as participation in a "historical opportunity" in large-language-model development, and structured 01.AI as an open-weights-first lab to establish credibility with developers outside China.

The Yi family's first significant release was Yi-34B in November 2023, an open-weights dense model that performed competitively against Llama 2 and established 01.AI's Hugging Face presence. Yi-34B was released under the permissive Yi license (allowing commercial use), and its Hugging Face leaderboard performance helped 01.AI reach a $1 billion valuation within eight months of founding. Smaller open-weights variants followed: Yi-6B and Yi-9B for edge and on-device deployment, plus the Yi-VL multimodal series (Yi-VL-34B and Yi-VL-6B) for vision-language tasks.

The Yi-1.5 series launched in May 2024, refining the base Yi architecture with improved instruction-following, coding, and mathematical reasoning. Yi-1.5 retained the open-weights distribution of its predecessors and extended the family to a broader parameter range.

Yi-Large arrived in June 2024 as the lab's first closed-source flagship, designed for use cases that require stronger performance than the open-weights Yi variants without the operational overhead of self-hosting. Where Yi-34B and Yi-1.5 were built for developer download and deployment, Yi-Large was positioned as an API-first enterprise offering. The model was trained on a large multilingual corpus with a knowledge cutoff of March 31, 2024.

Yi-Coder, a coding-assistant model covering 52 programming languages, shipped in September 2024. Yi-Lightning followed in October 2024 and displaced Yi-Large as the lab's flagship, offering an enhanced mixture-of-experts architecture, 40 percent faster inference, and a dramatically lower API price ($0.14 per million tokens versus Yi-Large's $3.00 per million).

The lab's trajectory changed substantially in December 2024, when 01.AI restructured its pre-training and infrastructure operations into a joint venture with Alibaba Cloud. Pre-training and infrastructure personnel transferred to Alibaba's Tongyi (Qwen) division; 01.AI retained a smaller application-focused team. In early 2025, 01.AI announced it would shift to an enterprise platform model -- Wanzhi -- powered in part by third-party frontier models including DeepSeek and Alibaba's Qwen, rather than continuing independent large-scale pre-training. Yi-Large thus represents the most capable dense model 01.AI developed before that strategic pivot.

Capabilities

Yi-Large handles multi-turn text conversation, document analysis, knowledge retrieval, data classification, and customer-service automation. The architecture is a decoder-only transformer with pre-normalization (applying layer normalization before the attention and feed-forward sub-layers for training stability), SwiGLU activation, Rotary Position Embedding (RoPE), and Group Query Attention (GQA) for efficient key-value caching at inference time.

The 32,768-token context window supports long-document and multi-turn session tasks without mid-session truncation. Multilingual proficiency is a design emphasis: in addition to English and Chinese, the model performs strongly across Spanish, Japanese, German, and French, reflecting the training corpus composition.

Yi-Large is not a reasoning or thinking model in the sense of extended chain-of-thought inference. It performs single-pass generation with no integrated scratchpad or reasoning trace. Mathematical and coding tasks are handled, but these are not the model's primary use-case framing; Yi-Lightning and Yi-Coder were built to address those more specifically.

The model's primary differentiators relative to the open-weights Yi family are better instruction-following fidelity and lower operational friction for API consumers who do not want to manage GPU infrastructure. Compared to Yi-34B self-hosted at equivalent parameter counts, Yi-Large offers stronger benchmark performance, which 01.AI attributed to additional training data, RLHF alignment, and post-training refinements not applied to the open-weights release.

Benchmarks and standing

Yi-Large's benchmark profile from the June 2024 release places it in the GPT-4-class range for general instruction-following and multilingual tasks, as characterized by 01.AI and by third-party evaluation at the time of launch. On Chatbot Arena, Yi-Large performed in the top tier for Chinese-language responses and showed competitive scores across math and hard-prompt categories.

On MMLU-Pro (a reasoning-intensive variant of the Massive Multitask Language Understanding benchmark), Yi-Large scored close to Claude 3 Sonnet at the time of evaluation. Independent third-party benchmark comparisons from mid-2024 placed Yi-Large roughly on par with GPT-3.5 and competitive with smaller Claude 3 models, though below GPT-4o and Claude 3 Opus in head-to-head evaluation.

As of April 2026, Yi-Large is not represented prominently on current standardized benchmark leaderboards such as the Artificial Analysis Intelligence Index or LMArena general ELO. The model's June 2024 release predates most of the evaluation infrastructure built around 2025 and 2026 frontier models. Yi-Lightning (October 2024), which superseded it, achieved a rank of 6th overall on Chatbot Arena at release and placed 2nd to 4th in specialized categories including Chinese, math, coding, and hard prompts -- the strongest recorded position for any Yi model on that leaderboard. Yi-Large's benchmark ceiling is lower than Yi-Lightning's.

Benchmark positions are point-in-time. The rapid cadence of Chinese model releases through 2024 and 2025 means Yi-Large's relative standing has declined since launch.

Access and pricing

Yi-Large is available through 01.AI's developer platform at platform.lingyiwanwu.com. Third-party inference providers include NVIDIA NIM at build.nvidia.com/01-ai/yi-large and OpenRouter at openrouter.ai/01-ai/yi-large. Fireworks AI and other mid-tier inference providers also carried Yi-Large at launch.

Pricing at third-party providers at the time of release was $3.00 per million input tokens and $3.00 per million output tokens, which was competitive with mid-tier US API offerings in mid-2024. The subsequent release of Yi-Lightning at $0.14 per million tokens on 01.AI's own platform sharply undercut Yi-Large on price. Availability and pricing across third-party providers as of April 2026 varies; some providers have reduced or discontinued Yi-Large access since Yi-Lightning's launch.

Yi-Large has no open-weights release. Source weights are not publicly distributed. For open-weights access to the Yi family, the most capable distributed variants are Yi-34B (November 2023), Yi-1.5-34B (May 2024), and the Yi-VL vision-language series, all available through the 01-ai organization on Hugging Face under the Yi license.

Comparison

Direct competitors to Yi-Large, considered as of the model's active deployment period (June to December 2024) and its continued API availability:

• DeepSeek V4 (DeepSeek). As of April 2026, DeepSeek V4 Pro is the most capable Chinese-origin model available for API access, with a 1.6-trillion-parameter MoE architecture, 1-million-token context window, and MIT-licensed open weights. Yi-Large was contemporary with DeepSeek V2 and pre-dates the V3/R1/V4 releases that transformed DeepSeek's global standing. On every capability axis relevant to April 2026, DeepSeek V4 Pro substantially exceeds Yi-Large.
• Qwen 3 (Alibaba Qwen). Qwen 3 is the current flagship from the Alibaba Qwen team, which includes the researchers transferred from 01.AI as part of the December 2024 restructuring. Qwen 3 operates at frontier scale and benchmarks competitively across multilingual tasks, coding, and reasoning -- domains where Yi-Large was also designed to perform. The relationship is notable: 01.AI's former pre-training team is now a structural contributor to Qwen's development.
• Kimi K2 (Moonshot AI). Kimi K2 is Moonshot AI's flagship model from 2025, competing at the high end of Chinese open-weights. Yi-Large predates Kimi K2 and does not compete with it on current leaderboards. Both emerged from the same 2023 Chinese AI startup cohort, with Moonshot AI continuing independent frontier-tier pre-training while 01.AI pivoted away from it.
• Yi-Lightning (01.AI). The closest comparison is Yi-Large's immediate successor within the same lab. Yi-Lightning's MoE architecture offers better performance on math, coding, and multilingual tasks, 40 percent faster inference, a top-6 Chatbot Arena ranking at launch, and a price 95 percent below Yi-Large at $0.14 per million tokens. For new API integrations in 2025 and 2026, Yi-Lightning is the more current choice; Yi-Large remains available but is not the recommended 01.AI model for new deployments.

Yi-Large's historical position: the first closed-source API offering from a credible Chinese AI lab in the mid-2024 period, and the highest-capability model 01.AI released under a purely proprietary serving arrangement before the December 2024 restructuring changed the lab's trajectory.

Outlook

Open questions for Yi-Large and 01.AI's model trajectory through 2026 and 2027:

• Continued API availability. Whether 01.AI maintains Yi-Large and Yi-Lightning API access as the lab's enterprise focus shifts further toward the Wanzhi platform and third-party model integration. If 01.AI deprioritizes API infrastructure, third-party providers may become the only access point.
• Yi family succession. Whether any Yi successor to Yi-Lightning ships under 01.AI's own brand, or whether the Alibaba Cloud joint venture produces future Yi-labeled models under shared branding with the Qwen team.
• Open-weights legacy. The Yi-34B and Yi-1.5 family remains in active use in fine-tuning and deployment workflows built before the December 2024 restructuring. The long-term maintenance of these weights -- and any future open-weights release from the restructured 01.AI -- is unresolved.
• 01.AI's application-tier execution. The Wanzhi enterprise platform is the post-restructuring commercial bet. Whether it generates revenue at a scale comparable to what a successful model API business might have produced is the central commercial question for the remaining 01.AI team.
• Regulatory and geopolitical context. Chinese-origin models face supply-chain and data-residency scrutiny from US federal and European regulatory bodies. The Yi-Large API's footprint in regulated Western markets was limited at launch and has not expanded significantly since.

Sources

• 01.AI: Yi-Large model card on NVIDIA NIM. Official model specifications, context window, and use-case framing at launch.
• OpenRouter: Yi Large API Pricing and Providers. Per-token pricing and provider availability data.
• Artificial Analysis: Yi-Large Performance and Price Analysis. Third-party intelligence and performance benchmark data.
• arXiv: Yi-Lightning Technical Report (2412.01253). Technical details for the Yi-Lightning successor model, including MoE architecture description and Chatbot Arena ranking.
• arXiv: Yi -- Open Foundation Models by 01.AI (2403.04652). 01.AI's foundational technical report covering the Yi model family architecture and base training.
• TechNode: Kaifu Lee's AI unicorn 01.AI restructures to focus on AI video editing and new ventures. February 2025 restructuring coverage.
• Rest of World: China's AI boom is driven by DeepSeek and chip restrictions. Context on 01.AI adopting DeepSeek models for its enterprise platform.
• Wikipedia: 01.AI. Company history, funding rounds, Yi family lineage.
• Hugging Face: 01-ai organization. Open-weights Yi model distribution, model cards, and release history.