ERNIE is the family of large language models from Baidu, a Chinese technology company that serves ERNIE models through its Qianfan model platform and the ERNIE Bot website, and publishes some ERNIE checkpoints as open weights on Hugging Face; its newest release, ERNIE 5.5 Flash Preview, appeared on 11 November 2026. This page lists the ERNIE models Baidu documents on its official pages, with version names, context windows, licences, published prices in yuan, release dates and status, plus an index of model pages and technical reports. Every fact comes from Baidu's ERNIE site, its Qianfan pricing documentation, its Hugging Face model cards and Baidu-issued press releases. Where Baidu publishes no value, the page says "Not stated by Baidu" rather than estimating. Facts are as of 15 November 2026.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- ERNIE 5.5 Flash Preview is the newest model. Released on 11 November 2026, it is a 280B total, 14B active Mixture-of-Experts model with a 1M-token context, served on Qianfan and listed on OpenRouter, according to Baidu's release page.
- ERNIE 5.1 is the current flagship on Qianfan. Baidu's ERNIE site records its official release on 9 May 2026, after a 30 April 2026 preview; Qianfan prices it in two input-length tiers up to 128K tokens.
- Prices are in yuan per thousand tokens. Qianfan's billing page, updated 11 November 2026, lists ERNIE 5.1 from CNY 0.004 input and 0.018 output per 1K tokens, and ERNIE 5.5 Flash Preview with peak and off-peak rates.
- Open weights exist for ERNIE 4.5 and a 5.5 Flash base checkpoint. The ten ERNIE 4.5 models are under Apache 2.0; ERNIE 5.5 Flash Base uses Baidu's own ERNIE 5.5 Flash Base Model License Agreement v1.0.
- No system cards are published. Baidu publishes release posts, Hugging Face model cards and an ERNIE 4.5 technical report, but no document labelled system card on the pages checked.
What is Baidu?
Baidu, Inc. is a Chinese technology company listed on NASDAQ (BIDU) and the Hong Kong Stock Exchange (9888), as stated in its press releases. Its ERNIE models are developed by the Baidu ERNIE team and trained with Baidu's PaddlePaddle deep learning framework. Baidu distributes them three ways: Qianfan (百度千帆), Baidu AI Cloud's Models-as-a-Service platform for enterprises and developers; ERNIE Bot at ernie.baidu.com and the Wenxiaoyan app for consumers; and Hugging Face and Baidu AI Studio for open-weight checkpoints. Baidu's ERNIE site at ernie.baidu.com carries release posts and research posts for each generation.
For other Chinese providers on this site, see Alibaba Qwen Models, DeepSeek Models, Moonshot AI Kimi Models and Zhipu GLM Models.
Current ERNIE models at a glance
As of 2026-11-15. Listed by release date, newest first. Sources: ERNIE site, Qianfan billing page and Hugging Face model cards.
| Model | API ID or version | Category | Context | Licence | Status |
|---|---|---|---|---|---|
| ERNIE 5.5 Flash Preview | ERNIE-5.5-Flash-Preview (Qianfan); baidu/ernie-5.5-flash-preview:free (OpenRouter link) | Post-trained text model | 1M tokens | Proprietary API; Baidu states weights will be released under Apache 2.0 | Preview |
| ERNIE 5.5 Flash Base | baidu/ERNIE-5.5-Flash-Base (Hugging Face) | Pre-trained base checkpoint | 1,048,576 for fine-tuning and inference | ERNIE 5.5 Flash Base Model License Agreement v1.0 | Released (open weights) |
| ERNIE 5.1 | ERNIE-5.1 (Qianfan) | Flagship model | Priced up to 128K input | No open weights listed | Current |
| ERNIE 5.0 | Not stated by Baidu on the pages checked (2026-11-15) | Natively omni-modal foundation model | Not stated by Baidu on the pages checked (2026-11-15) | No open weights listed | Preview at announcement (Nov 2025) |
| ERNIE X1.1 | Not stated by Baidu on the pages checked (2026-11-15) | Reasoning model | Not stated by Baidu on the pages checked (2026-11-15) | No open weights listed | Released Sep 2025 |
| ERNIE-4.5-21B-A3B-Thinking | Hugging Face and Baidu AI Studio | Open thinking model | 128K | Apache 2.0 | Released (open weights) |
| ERNIE 4.5 family (10 models) | e.g. ERNIE-4.5-300B-A47B-PT, ERNIE-4.5-VL-28B-A3B-PT | Open text and vision-language models | 131,072 (300B-A47B card) | Apache 2.0 | Released (open weights) |
The Qianfan billing page also lists ERNIE-Light-DT-1.0 and older fine-tunable ERNIE Speed and ERNIE Lite versions; they are covered under pricing.
Key facts for the main ERNIE models
As of 2026-11-15.
| Model | Released | Max output | Knowledge cutoff | Input / output | Open weights | Parameters | Positioning (provider wording) |
|---|---|---|---|---|---|---|---|
| ERNIE 5.5 Flash Preview | 2026-11-11 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Text / text | Not yet (Apache 2.0 stated as planned) | 280B total, 14B active | Coding, 3D modelling, knowledge work and content creation; "an intermediate checkpoint from post-training" |
| ERNIE 5.5 Flash Base | 2026-11-05 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Text / text | Yes (FP8 and BF16) | 280B total, 14B active | End-of-pre-training checkpoint for domain adaptation, instruction tuning and RL research |
| ERNIE 5.1 | 2026-05-09 (preview 2026-04-30) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | No | Not stated by Baidu on the pages checked (2026-11-15) | "A Model That Writes Better and Understands You More" |
| ERNIE 5.0 | 2025-11-13 (announced, preview) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Text, image, audio, video (understanding and generation) | No | Not stated by Baidu on the pages checked (2026-11-15) | Natively omni-modal foundation model |
| ERNIE X1.1 | 2025-09-09 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | No | Not stated by Baidu on the pages checked (2026-11-15) | Reasoning model built on ERNIE 4.5 |
| ERNIE 4.5 (300B-A47B) | 2025-06-30 (open-source release) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Text / text | Yes | 300B total, 47B active | Largest text MoE in the ERNIE 4.5 family |
ERNIE model families and naming
- Numbered generations (4.5, 5.0, 5.1, 5.5): foundation model generations. Point releases such as 5.1 follow a preview, and Baidu labels early checkpoints "Preview".
- X series (X1, X1.1): reasoning models. Baidu's press release states that X1.1 is built on ERNIE 4.5 and trained with an iterative hybrid reinforcement learning framework.
- Flash: the 5.5 Flash line is a sparse Mixture-of-Experts design with 14B active parameters. "Base" marks the pre-trained checkpoint; "Preview" marks the first post-trained model built on it.
- Open-weight naming (ERNIE 4.5): total and active parameters appear in the name, for example 300B-A47B and 21B-A3B. "VL" marks vision-language models, "PT" marks PyTorch weights, "Paddle" marks PaddlePaddle weights, and "Thinking" marks the deep-thinking variant.
- Qianfan version names: the API uses version strings such as ERNIE-5.1 and ERNIE-5.5-Flash-Preview; older lines (ERNIE Speed, ERNIE Lite) carry context size and date suffixes such as ERNIE-Lite-8K-0308.
When was each ERNIE model released?
As of 2026-11-15. Dates as stated on Baidu's ERNIE site or in Baidu-issued press releases.
| Date | Event | API ID |
|---|---|---|
| 2025-06-30 | Ten ERNIE 4.5 models open-sourced under Apache 2.0 | e.g. ERNIE-4.5-300B-A47B, ERNIE-4.5-21B-A3B, ERNIE-4.5-VL-424B-A47B, ERNIE-4.5-0.3B |
| 2025-09-09 | ERNIE X1.1 announced at WAVE SUMMIT 2025; ERNIE-4.5-21B-A3B-Thinking open-sourced | Not stated by Baidu on the pages checked (2026-11-15) |
| 2025-11-13 | ERNIE 5.0 announced at Baidu World 2025; preview on ERNIE Bot and Qianfan | Not stated by Baidu on the pages checked (2026-11-15) |
| 2026-04-30 | ERNIE 5.1-Preview posted to LMArena | Not stated by Baidu on the pages checked (2026-11-15) |
| 2026-05-09 | ERNIE 5.1 officially released | ERNIE-5.1 |
| 2026-11-05 | ERNIE 5.5 Flash Base released on Hugging Face | baidu/ERNIE-5.5-Flash-Base |
| 2026-11-11 | ERNIE 5.5 Flash Preview released on Qianfan and OpenRouter | ERNIE-5.5-Flash-Preview |
The X1.1 press release reports, against ERNIE X1, gains of 34.8% in factuality, 12.5% in instruction following and 9.6% in agentic capability. Baidu does not give the X1 release date on the pages checked.
How much do ERNIE models cost?
As of 2026-11-15. CNY (yuan) per 1,000 tokens, online inference, pay as you go. Source: Qianfan model service billing page (模型服务计费), updated 2026-11-11.
| Model | Input | Cached input | Output | Batch | Cache write | Notes |
|---|---|---|---|---|---|---|
| ERNIE 5.5 Flash Preview, peak (08:00 to 22:00) | 0.001 promo (list 0.002) | 0.0001 promo (list 0.0002) | 0.004 promo (list 0.008) | Not priced | Not stated by Baidu on the pages checked (2026-11-15) | Promotional prices shown alongside list prices |
| ERNIE 5.5 Flash Preview, off-peak (22:00 to 08:00) | 0.0005 promo (list 0.001) | 0.00005 promo (list 0.0001) | 0.002 promo (list 0.004) | Not priced | Not stated by Baidu on the pages checked (2026-11-15) | Half the peak rate |
| ERNIE 5.1, input up to 32K | 0.004 | Not stated by Baidu on the pages checked (2026-11-15) | 0.018 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Tiered by input length |
| ERNIE 5.1, input 32K to 128K | 0.006 | Not stated by Baidu on the pages checked (2026-11-15) | 0.022 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Tiered by input length |
| ERNIE-Light-DT-1.0 | 0.008 | 0.0008 | 0.028 | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | No peak and off-peak split |
| ERNIE-Speed-8K (fine-tuned, per token) | 0.004 discounted (list 0.012) | Not stated by Baidu on the pages checked (2026-11-15) | 0.008 discounted (list 0.024) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Small-traffic trials; one such service per account |
| ERNIE-Lite-8K-0308 (fine-tuned, per token) | 0.003 discounted (list 0.009) | Not stated by Baidu on the pages checked (2026-11-15) | 0.006 discounted (list 0.018) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Other ERNIE Lite versions list a single 0.008 rate |
- Per-million conversion: multiply by 1,000. ERNIE 5.1 at up to 32K input is CNY 4 input and CNY 18 output per million tokens.
- Worked example: an ERNIE 5.1 request with 20,000 input and 2,000 output tokens costs 20 x 0.004 plus 2 x 0.018, which is CNY 0.116.
- Open weights: self-hosting ERNIE 4.5 or 5.5 Flash Base carries no Baidu token charge; infrastructure costs fall on the operator.
- OpenRouter: Baidu's release page links an OpenRouter listing with a free tier for ERNIE 5.5 Flash Preview; OpenRouter's own terms apply.
For prices across providers see Frontier AI Models Compared.
Where are ERNIE models available?
| Channel | Details | Source |
|---|---|---|
| Baidu Qianfan | ERNIE 5.5 Flash Preview, ERNIE 5.1, ERNIE-Light-DT, ERNIE Speed and ERNIE Lite; ERNIE X1.1 and ERNIE 5.0 at their launch | Qianfan billing page; Baidu press releases |
| ERNIE Bot (ernie.baidu.com) and Wenxiaoyan app | ERNIE X1.1 and ERNIE 5.0 preview at launch | Baidu press releases |
| OpenRouter | ERNIE 5.5 Flash Preview, with a free tier listing | ERNIE 5.5 Flash Preview release page |
| Hugging Face | ERNIE 4.5 family, ERNIE-4.5-21B-A3B-Thinking, ERNIE 5.5 Flash Base (FP8 and BF16) | Hugging Face model cards; ERNIE 4.5 post |
| Baidu AI Studio | ERNIE-4.5-21B-A3B-Thinking | X1.1 press release |
| Self-hosted serving | ERNIE 5.5 Flash Base via SGLang and vLLM; ERNIE 4.5 via FastDeploy and ERNIEKit | Hugging Face model card; ERNIE 4.5 post |
| Other clouds | Not stated by Baidu on the pages checked (2026-11-15) | None |
Restricted-access ERNIE models
Not stated by Baidu on the pages checked (2026-11-15): Baidu publishes no allowlist-only or trusted-access ERNIE model on the pages checked. The access conditions it does state are below.
| Model | Access stated by provider | Data retention condition |
|---|---|---|
| ERNIE 5.5 Flash Preview | Public preview on Qianfan; Baidu calls it an intermediate post-training checkpoint | Not stated by Baidu on the pages checked (2026-11-15) |
| Fine-tuned ERNIE Speed and Lite (per token) | One pay-per-token fine-tuned service per Baidu Intelligent Cloud account; publishing fails when capacity is full | Not stated by Baidu on the pages checked (2026-11-15) |
What is the context window of ERNIE models?
- ERNIE 5.5 Flash Preview: 1M tokens (release page).
- ERNIE 5.5 Flash Base: trained at up to 131,072 tokens; the released configuration enables 1,048,576 tokens for fine-tuning and inference.
- ERNIE 5.1: Qianfan prices input in tiers up to 128K tokens; Baidu does not state a separate context window figure on the pages checked.
- ERNIE-4.5-300B-A47B: 131,072 tokens (Hugging Face model card).
- ERNIE-4.5-21B-A3B-Thinking: 128K tokens (X1.1 press release).
- ERNIE 5.0 and ERNIE X1.1: Not stated by Baidu on the pages checked (2026-11-15).
See Context Windows Explained.
Is any ERNIE model open source?
Yes, some. Baidu's ERNIE 4.5 post states that "All models are publicly accessible under Apache 2.0", covering ten ERNIE 4.5 models released on 30 June 2025, and that the licence permits commercial use. ERNIE-4.5-21B-A3B-Thinking followed in September 2025. ERNIE 5.5 Flash Base is open weight under the ERNIE 5.5 Flash Base Model License Agreement v1.0, a Baidu licence rather than an OSI-approved one; read the LICENSE file before use. For ERNIE 5.5 Flash Preview, Baidu states the weights "will be open-sourced under Apache 2.0 soon". ERNIE 5.1, 5.0 and X1.1 are served through Baidu's own channels with no open weights on the pages checked.
See Open-Weight vs Closed Models Explained and the Open-Weight Model Security Guide.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| ERNIE 5.5 Flash Preview | Release page on ernie.baidu.com | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) |
| ERNIE 5.5 Flash Base | Hugging Face model card; release page | Not stated by Baidu on the pages checked (2026-11-15) | Component papers on arXiv (Virtual Head Attention, Multi-View MoE, ECHO) |
| ERNIE 5.1 | Release post on ernie.baidu.com | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) |
| ERNIE 5.0 | Research post on ernie.baidu.com | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) |
| ERNIE 4.5 | Hugging Face model cards; research post | Not stated by Baidu on the pages checked (2026-11-15) | ERNIE 4.5 Technical Report (Baidu ERNIE site) |
The ERNIE 5.5 Flash Base card documents 43 layers (1 dense and 42 MoE), 512 routed experts with top-10 routing plus a shared expert, a 201,216-entry vocabulary, 33.5T pre-training tokens with the Muon optimiser and FP8 mixed precision. See Model Cards and System Cards Explained.
Benchmarks
Provider-reported scores, from Baidu's Hugging Face model cards, as shown on 2026-11-15. ERNIE 5.5 Flash Base scores are base-model results with the shot counts stated in the card; the Flash Preview release page shows its scores only as images, so they are not reproduced. See How to Read LLM Benchmarks.
| Benchmark | ERNIE 5.5 Flash Base (provider-reported) | ERNIE-4.5-300B-A47B (provider-reported) | Independent leaderboard results |
|---|---|---|---|
| MMLU-Pro (5-shot) | 78.4 | 78.4 | Not checked for this page |
| MMLU (5-shot) | 93.3 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
| GPQA (5-shot) | 60.2 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
| CMMLU (5-shot) | 90.0 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
| HumanEval+ (1-shot) | 82.3 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
| LiveCodeBench (1-shot) | 40.1 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
| LongBench-v2 (0-shot) | 48.7 | Not stated by Baidu on the pages checked (2026-11-15) | Not checked for this page |
Baidu's ERNIE site records that ERNIE 5.1-Preview ranked as the top Chinese model on the LMArena Text Leaderboard in a post dated 30 April 2026; that is Baidu's statement, not a reading of the leaderboard for this page. The X1.1 press release states that X1.1 is on par with GPT-5 and Gemini 2.5 Pro; Baidu does not publish the scores behind that statement on the pages checked.
Deprecations and retirements
| Model | API ID | Deprecated or notice | Retirement or shutdown | Stated replacement |
|---|---|---|---|---|
| All ERNIE models listed | As above | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) | Not stated by Baidu on the pages checked (2026-11-15) |
The Qianfan billing page still prices dated ERNIE Lite versions (0308, 0516, 0704, 0725, 0922 suffixes), which shows older versions stay listed alongside new ones. Check Qianfan's documentation for lifecycle notices before building on any version. See the AI Model Retirement Guide.
Data handling and security
- Qianfan API data use and retention: Not stated by Baidu on the pages checked (2026-11-15). Read Baidu AI Cloud's service agreement and privacy terms before sending personal or confidential data.
- Self-hosted open weights: prompts and outputs stay on the operator's infrastructure; the operator owns access control, logging and patching.
- Jurisdiction: Qianfan is a Chinese cloud service. China's generative AI measures and data laws apply to services offered in China; see the China Generative AI Measures and AI Labelling Rules Guide and China Cross-Border Data Transfers.
- Third-party routes: traffic through OpenRouter follows OpenRouter's data terms as well as the provider's.
- Credentials: Qianfan API keys are non-human identities with spend and data exposure; see the LLMjacking Guide.
How to choose an ERNIE model: factual criteria
| Requirement | Documented fact to check |
|---|---|
| Weights you can host and modify | ERNIE 4.5 (Apache 2.0) or ERNIE 5.5 Flash Base (Baidu licence v1.0) |
| Inputs above 128K tokens | ERNIE 5.5 Flash Preview states 1M; ERNIE 5.1 pricing tiers stop at 128K |
| Lowest published Qianfan token price | ERNIE 5.5 Flash Preview off-peak promotional rates |
| Generally released rather than preview | ERNIE 5.1 is officially released; 5.5 Flash Preview is an intermediate checkpoint |
| Image, audio or video input | ERNIE 5.0 (omni-modal) or ERNIE 4.5 VL models; 5.5 Flash is text only |
| Explicit reasoning model | ERNIE X1.1 or ERNIE-4.5-21B-A3B-Thinking |
Evaluate shortlisted models on your own data; provider benchmark tables use different shot counts and comparators.
Practitioner checklist
- Record the exact Qianfan version string (for example ERNIE-5.1) and the date adopted.
- Budget in yuan per thousand tokens, and check whether promotional or list prices apply on the billing date.
- Schedule batch-like workloads off-peak (22:00 to 08:00) where ERNIE 5.5 Flash Preview rates halve.
- Treat ERNIE 5.5 Flash Preview as a preview checkpoint: pin versions and retest when Baidu updates it.
- Read the ERNIE 5.5 Flash Base licence before commercial fine-tuning; it is not Apache 2.0.
- Confirm Baidu AI Cloud data terms and cross-border data rules before sending personal data to Qianfan.
- Store Qianfan API keys in a secrets manager and set spend alerts.
- Run your own evaluations; provider-reported scores are not independent.
Standards and references
- Baidu AI Cloud: Qianfan model service billing (accessed 2026-11-15)
- Baidu: ERNIE site (accessed 2026-11-15)
- Baidu: ERNIE 5.5 Flash Preview release (accessed 2026-11-15)
- Baidu: ERNIE 5.5 Flash Base release (accessed 2026-11-15)
- Hugging Face: baidu/ERNIE-5.5-Flash-Base model card (accessed 2026-11-15)
- Baidu: ERNIE 4.5 open-source release (accessed 2026-11-15)
- Hugging Face: baidu/ERNIE-4.5-300B-A47B-PT model card (accessed 2026-11-15)
- Baidu press release: ERNIE X1.1 (accessed 2026-11-15)
- Baidu press release: ERNIE 5.0 at Baidu World 2025 (accessed 2026-11-15)
Related NHI Mgmt Group resources: Alibaba Qwen Models, DeepSeek Models, Frontier AI Models Compared, Open-Weight Model Security Guide.