ERNIE is Baidu's flagship large-language-model family and the longest-running Chinese frontier-model line. The general-purpose ERNIE 4 (October 2023) and the reasoning-tier ERNIE X1 (2025) are the closed-weight flagship models behind Wenxin Yiyan (ERNIE Bot, 文心一言), which launched on 16 March 2023 as the first Chinese ChatGPT-equivalent public service.
| Field | Value | |
|---|---|---|
| Developer | [[baidu-ernie | Baidu]] |
| Model family name | ERNIE (Enhanced Representation through Knowledge Integration) / 文心 (Wenxin) | |
| Lineage | ERNIE 1.0 (2019) → ERNIE 3.0 / Titan (2021) → ERNIE Bot (Mar 2023) → ERNIE 4 (Oct 2023) → ERNIE 4 Turbo (2024) → ERNIE X1 (2025) | |
| Deployment | Wenxin Yiyan (ERNIE Bot) consumer chatbot; Baidu AI Cloud APIs | |
| License | Closed-weight for flagship; smaller ERNIE-Speed / ERNIE-Lite / ERNIE-Tiny variants partially open | |
| Consumer pricing | Free to consumers since 2024 (Source: China and the US Are Running Different AI Races) |
ERNIE Bot launched publicly on 16 March 2023, roughly four months after ChatGPT's November 2022 launch, and was the first Chinese ChatGPT-equivalent public service. ERNIE 4 followed in October 2023, with Baidu claiming parity with GPT-4 on Chinese benchmarks. ERNIE X1 (2025) is Baidu's reasoning-tier entry, positioned against DeepSeek-R1 and OpenAI's o-series. The ERNIE lineage predates the modern LLM era: ERNIE 1.0 (2019) was a BERT-style knowledge-enhanced encoder, making Baidu one of the few companies with a continuous deep-learning research program spanning the pre-GPT and post-GPT eras. ERNIE Bot's March 2023 launch is cited in US-China AI competition discussion as the point at which China shipped a ChatGPT equivalent within months of OpenAI.
Model variants
| Model | Release | Notes |
|---|---|---|
| ERNIE 1.0 | 2019 | BERT-era knowledge-enhanced encoder |
| ERNIE 3.0 / ERNIE 3.0 Titan | 2021 | 260B-parameter precursor |
| ERNIE 3.5 | 2023 | Underlying model at ERNIE Bot launch |
| ERNIE 4 | Oct 2023 | Baidu's "GPT-4-class" flagship; closed-weight |
| ERNIE 4 Turbo | 2024 | Efficiency / latency variant |
| ERNIE-Speed / ERNIE-Lite / ERNIE-Tiny | 2024 | Smaller tiers; some partially open |
| ERNIE X1 | 2025 | Reasoning-tier model |
Capabilities
ERNIE is optimized for Chinese-first use cases, and Baidu-reported benchmarks routinely show ERNIE 4 at or above GPT-4 on Chinese-language tasks, with weaker results on English benchmarks. These figures are Baidu's own numbers and have not been independently replicated (confidence: medium). ERNIE's historical differentiator was integration with Baidu's knowledge graph; how much of that persists in the ERNIE 4 architecture is unclear from public disclosures. ERNIE 4 supports image understanding and generation, integrated into the Wenxin Yiyan consumer product. ERNIE X1 is Baidu's response to the reasoning-model trend and is deployed separately from the ERNIE 4 general-purpose model; it is another instance of Chinese labs shipping reasoning-tier models within months of o1 and DeepSeek-R1, as discussed in the fast-follow problem.
Open-weight status
ERNIE 4 and ERNIE X1 are not open-weight. Smaller variants (ERNIE-Speed-8K, ERNIE-Lite) are accessible via API, with partial open releases at small scales. Baidu CEO Robin Li has publicly argued that open-sourcing frontier models is "not a good idea," a stance that runs counter to those of DeepSeek and Alibaba, which have pursued open-weight strategies (confidence: medium on quote attribution).
ERNIE serves as a Chinese counterexample to the generalization that Chinese labs ship open-weight models while US labs ship closed, a pattern associated with DeepSeek-V3 and Qwen3. This counterexample is the basis for the page's contradicts-relationship to open-source AI.
Availability, pricing, and scale
ERNIE Bot has been free to consumers since 2024, with enterprise monetization via Baidu AI Cloud APIs. This reflects a Chinese pattern of monetizing through cloud, API, and advertising rather than consumer subscriptions, as discussed in China and the US Are Running Different AI Races. Baidu reports 200M+ ERNIE Bot users at various points; these are marketing claims (confidence: low).
Regulatory environment
ERNIE Bot and the underlying models operate under several Chinese frameworks. Under the China — Interim Measures for the Management of Generative AI Services (Aug 2023), Baidu was among the first batch registered on the effective date. The China — Provisions on the Administration of Deep Synthesis Internet Information Services (Jan 2023) apply to ERNIE's image and video outputs, and the China — Internet Information Service Algorithmic Recommendation Management Provisions (Mar 2022) cover Baidu Search surfaces that use ERNIE. Oversight is by the CAC. Content alignment requirements are substantive and shape ERNIE Bot's refusal behavior on politically sensitive topics.
Comparison to peer models
| Model | Developer | Open-weight | Reasoning mode | Distinctive | ||
|---|---|---|---|---|---|---|
| ERNIE 4 / X1 | Baidu | No | Separate X1 model | First public Chinese ChatGPT-equivalent; closed | ||
| [[qwen3 | Qwen3]] | Alibaba | Apache 2.0 at 235B | Unified dual-mode | Most open flagship | |
| [[deepseek-v3 | DeepSeek-V3]] / [[deepseek-r1 | R1]] | DeepSeek | Fully open | Separate R1 model | Efficiency focus |
| [[glm-4 | GLM-4 / 4.5]] | Zhipu | Partial (9B open) | Unified in 4.5 | Tsinghua pedigree | |
| [[kimi-k2 | Kimi K2]] | Moonshot | Research-only license | — | Long-context / agentic | |
| [[doubao | Doubao]] | ByteDance | Largely closed | Variants reported | Consumer-scale distribution |
Relationships
- instance-of: General-Purpose AI (GPAI)
- related: DeepSeek-V3, DeepSeek-R1, Qwen3, GLM-4 generation (GLM-4 through GLM-4.7), Kimi K2, Doubao, China and the US Are Running Different AI Races, Open-Source AI / Open-Weight Models
- contradicts: "Chinese labs ship open-weight" generalization on Open-Source AI / Open-Weight Models.
- depends-on: Scaling Laws
Sources
Referenced via Baidu (ERNIE) and China and the US Are Running Different AI Races. Confidence is medium: Baidu publishes fewer frontier technical papers than DeepSeek or Alibaba, and most claims rely on press reporting and Baidu's own disclosures.