ERNIE 5.1 is a mixture-of-experts large language model from Baidu, officially released on 9 May 2026 after a preview on 30 April 2026, which Baidu's release post headlines as "A Model That Writes Better and Understands You More". This profile gathers the published facts about it in one place: how it was derived from ERNIE 5.0, Qianfan prices, where it can be used, the leaderboard positions and scores Baidu reports, and what Baidu does not state. Every fact comes from Baidu's ERNIE site, the Baidu AI Cloud Qianfan billing page or a Baidu press release, listed under Standards and references. Where Baidu publishes nothing on a point, the page says so rather than estimating.
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. It uses Baidu's public ERNIE documentation as of November 2026. Model details change, so confirm against Baidu's current documentation.
Key takeaways
- Released 9 May 2026. Baidu's ERNIE site records ERNIE 5.1's official release on that date, after ERNIE-5.1-Preview appeared on the LMArena Text leaderboard on 30 April 2026.
- A smaller sub-network of ERNIE 5.0. Baidu states that ERNIE 5.1 inherits ERNIE 5.0's pre-training, with total parameters compressed to about one third and active parameters to about one half, and pre-training compute about 6% of comparable models.
- CNY 0.004 input and 0.018 output per 1K tokens. Qianfan's billing page, updated 11 November 2026, prices ERNIE 5.1 in two input-length tiers, up to 32K and 32K to 128K tokens.
- No published weights. ERNIE 5.1 is served through Baidu's own channels; no open weights or licence file appear on the pages checked.
- Many specifications not stated. Baidu does not state a context window, maximum output, modalities, knowledge cutoff, system card or deprecation date for ERNIE 5.1 on the pages checked.
What is ERNIE 5.1?
ERNIE 5.1 is the point release that followed ERNIE 5.0, which Baidu unveiled on 13 November 2025 at Baidu World 2025 as a "natively omni-modal" model for text, images, audio and video. Rather than training ERNIE 5.1 from scratch, Baidu derived it from a sub-network of ERNIE 5.0 using what its release post calls a "Once-For-All" elastic training framework, which trains many sub-models in one run.
The elasticity covers three dimensions, according to Baidu: depth (the number of active Transformer layers), width or expert capacity (the number of mixture-of-experts experts available to routing) and sparsity (a variable Top-k for the number of activated experts). Baidu states the result has significantly lower inference cost than ERNIE 5.0. It does not publish absolute parameter counts for either model on the pages checked, so only the ratios can be given.
Baidu positions ERNIE 5.1 for writing and creative work as well as reasoning and agentic tasks: the release post emphasises intent understanding, long-form narrative control and professional content. For the full ERNIE line-up, see Baidu ERNIE Models. For how sparse expert models work, see Mixture of Experts Explained.
Key facts at a glance
As of 2026-11-15. Sources: ERNIE 5.1 release post, ERNIE 5.1-Preview post and Qianfan billing page.
| Field | Value | Source |
|---|---|---|
| Provider | Baidu | ERNIE 5.1 release post |
| API ID | ERNIE-5.1 (Qianfan model version name) | Qianfan billing page |
| Release date | 2026-05-09 (preview 2026-04-30) | ERNIE 5.1 release post; preview post |
| Status | Current, officially released | ERNIE 5.1 release post |
| Context window | Not stated by Baidu on the pages checked (2026-11-15); Qianfan pricing tiers run to 128K input tokens | Qianfan billing page |
| Max output | Not stated by Baidu on the pages checked (2026-11-15) | ERNIE 5.1 release post |
| Knowledge cutoff | Not stated by Baidu on the pages checked (2026-11-15) | ERNIE 5.1 release post |
| Input modalities | Not stated by Baidu on the pages checked (2026-11-15); the release post describes text tasks | ERNIE 5.1 release post |
| Output modalities | Not stated by Baidu on the pages checked (2026-11-15) | ERNIE 5.1 release post |
| Architecture | Mixture of experts, elastic sub-network of ERNIE 5.0 | ERNIE 5.1 release post |
| Parameters | About one third of ERNIE 5.0 total and one half of its active parameters; absolute counts not stated by Baidu on the pages checked (2026-11-15) | ERNIE 5.1 release post |
| Open weights | No weights published on the pages checked | ERNIE site; Qianfan billing page |
| Licence | Not stated by Baidu on the pages checked (2026-11-15); no model licence file published | ERNIE site |
| Training framework | PaddlePaddle (asynchronous reinforcement learning system) | ERNIE 5.1 release post |
Release and version history
As of 2026-11-15. Dates from Baidu's ERNIE site, Qianfan billing page and Baidu press release.
| Date | Event | Source |
|---|---|---|
| 2025-11-13 | ERNIE 5.0 unveiled at Baidu World 2025 as a preview, available through ERNIE Bot and Qianfan | Baidu press release |
| 2026-04-30 | ERNIE-5.1-Preview appears on the LMArena Text leaderboard | ERNIE 5.1-Preview post |
| 2026-05-09 | ERNIE 5.1 officially released; Arena Search position reported on the same date | ERNIE 5.1 release post |
| 2026-11-11 | Latest update to the Qianfan billing page listing ERNIE 5.1 prices | Qianfan billing page |
Baidu labels early checkpoints "Preview" and then publishes the official release; ERNIE 5.1 followed that pattern, with nine days between the two. The ERNIE site also lists the later ERNIE 5.5 Flash Preview and ERNIE 5.5 Flash Base. Qianfan lists one version name, ERNIE-5.1, with no dated snapshots on the pages checked.
How much does ERNIE 5.1 cost?
As of 2026-11-15. CNY per 1,000 tokens, pay-as-you-go online inference. Source: Baidu AI Cloud Qianfan billing page (updated 2026-11-11).
| Item | Price | Notes | Source |
|---|---|---|---|
| Input up to 32K tokens | Input: CNY 0.004; Output: CNY 0.018 | CNY 4 and CNY 18 per million tokens | Qianfan billing page |
| Input 32K to 128K tokens | Input: CNY 0.006; Output: CNY 0.022 | CNY 6 and CNY 22 per million tokens | Qianfan billing page |
| Cached input | Not stated by Baidu on the pages checked (2026-11-15) | No cache-hit rate listed | Qianfan billing page |
| Batch inference | Not offered for ERNIE 5.1 in the billing table | Marked "-" | Qianfan billing page |
| Consumer chat | Not stated by Baidu on the pages checked (2026-11-15) | ernie.baidu.com chat | ERNIE 5.1 release post |
- Tier rule: the tier is set by the request's input length, and the higher tier raises both the input and output rate.
- Currency: Baidu prices in yuan. Convert at the rate on your invoice date rather than a fixed rate.
- Units: Qianfan quotes per 1,000 tokens; most other providers quote per million, so multiply by 1,000 before comparing.
Worked example in the lower tier: a request with 20,000 input tokens and 2,000 output tokens costs 20 x 0.004 plus 2 x 0.018, which is CNY 0.116. A request with 100,000 input and 10,000 output tokens falls in the upper tier: 100 x 0.006 plus 10 x 0.022, which is CNY 0.82. See Model Pricing Explained and the LLM Pricing Tracker.
Context window and output limits
Baidu does not state a context window or maximum output for ERNIE 5.1 on the pages checked. The only published limit is the pricing structure: Qianfan prices input up to 128K tokens, so 128K is the largest input length with a published price. Treat it as a billing boundary, not a confirmed context specification, and confirm the limit in the Qianfan console before designing long-context workloads.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | Not stated by Baidu on the pages checked (2026-11-15) | Pricing tiers run to 128K input tokens | Qianfan billing page |
| Max output | Not stated by Baidu on the pages checked (2026-11-15) | None | Qianfan billing page |
| Pricing tier boundary | 32,000 input tokens | Higher input and output rates above it | Qianfan billing page |
| Highest priced input length | 128,000 tokens | Upper tier covers 32K to 128K | Qianfan billing page |
| Rate limits | Not stated by Baidu on the pages checked (2026-11-15) | None | Qianfan billing page |
See Context Windows Explained.
Modalities and tools
Baidu does not list ERNIE 5.1's input and output modalities on the pages checked. ERNIE 5.0, its parent, is described as natively omni-modal, but Baidu does not say whether ERNIE 5.1 keeps that capability, and the 5.1 release post describes text tasks: writing, reasoning, tool use and agentic work.
- Tool use: the release post says the first post-training stage covers instruction following and tool invocation, and that domain experts were trained for code, reasoning and agentic tasks.
- Reported agentic evaluations: Baidu names tau3-bench and SpreadsheetBench-Verified and an AIME26 score with tool use (see Benchmarks).
- Reasoning controls: Not stated by Baidu on the pages checked (2026-11-15).
- Structured output and function calling parameters: Not stated by Baidu for ERNIE 5.1 on the pages checked (2026-11-15).
Where ERNIE 5.1 is connected to tools through an agent platform, each connection is a credentialed integration; see the LLM Security Guide.
Where is ERNIE 5.1 available?
As of 2026-11-15. Sources: ERNIE 5.1 release post and Qianfan billing page.
| Channel | Details | Source |
|---|---|---|
| Baidu AI Cloud Qianfan | Version name ERNIE-5.1; pay-as-you-go online inference | Qianfan billing page |
| ERNIE chat | Official chat site at ernie.baidu.com | ERNIE 5.1 release post |
| Baidu AI Studio | ERNIE 5.1 Playground | ERNIE 5.1 release post |
| Creative agent platforms | Rollout to more than ten platforms; named examples ISEKAI ZERO, Mulan AI, Diting Huanliu and Storymaster | ERNIE 5.1 release post |
| Hyperscaler clouds outside Baidu | Not stated by Baidu on the pages checked (2026-11-15) | ERNIE 5.1 release post |
| Downloadable weights | Not published on the pages checked | ERNIE site |
Region availability and account requirements for Qianfan are not stated on the pages checked. Third-party creative platforms run their own services under their own terms.
Access restrictions
ERNIE 5.1 is not a restricted-access model: it is priced for pay-as-you-go use on Qianfan and offered in Baidu's public chat and AI Studio playground. The restrictions that apply are those of the service rather than a model licence. Because no weights are published, ERNIE 5.1 cannot be self-hosted; use depends on Baidu's channels and their terms. Read the Baidu AI Cloud service agreement before commercial use. See Open-Weight vs Closed LLMs Explained.
Benchmarks (provider-reported, dated)
As of 2026-11-15. All entries are provider-reported, from Baidu's ERNIE 5.1 release post (9 May 2026) and ERNIE 5.1-Preview post (30 April 2026). Leaderboard positions are Baidu's statements about the leaderboards, not readings of the leaderboards for this page. Baidu's benchmark chart is an image; only figures stated in text are listed.
| Evaluation | ERNIE 5.1 | Comparator in source | Source and date |
|---|---|---|---|
| LMArena Text leaderboard (Preview) | 13th globally, 1st among Chinese models | None named | Preview post, 2026-04-30 |
| LMArena Text categories (Preview) | Legal and Government 1st; Business, Management and Financial Ops 4th; Software and IT Services 7th; Math 9th | None named | Preview post, 2026-04-30 |
| Arena Search leaderboard | 1,223 points; 4th globally, 1st among Chinese models | None named | Release post, scored 2026-05-09 |
| AIME26 (with tool use) | 99.6 | "Second only to Gemini 3.1 Pro" | Release post, 2026-05-09 |
| tau3-bench, SpreadsheetBench-Verified | Scores not stated in text | Baidu states ERNIE 5.1 surpasses DeepSeek-V4-Pro | Release post, 2026-05-09 |
| GPQA, MMLU-Pro | Scores not stated in text | Baidu describes results as approaching leading closed-source models | Release post, 2026-05-09 |
Leaderboard positions shift as models are added, so a rank is only meaningful with its date. Run your own evaluation before relying on rank claims. See How to Read LLM Benchmarks and the LLM Leaderboard Snapshot.
Model card and system card
Baidu does not publish a model card or system card for ERNIE 5.1 on the pages checked. The release post is the main technical source, and it covers training methods in some depth:
- Elastic pre-training: a sub-network of ERNIE 5.0 selected with the Once-For-All framework across depth, expert width and sparsity.
- Reinforcement learning system: a disaggregated, fully asynchronous system on PaddlePaddle, with an RL controller separating training, inference, reward and agent-loop subsystems; FP8 training-inference consistency work; and an optimised Rollout Router Replay technique that Baidu says halves a KL divergence measure with near-zero added latency.
- Post-training in four stages: unified supervised fine-tuning, parallel domain expert training (code, reasoning, agentic), on-policy distillation with the experts as teachers using token-level reverse KL, then general online reinforcement learning for conversational and creative tasks.
- Safety evaluations: Not stated by Baidu on the pages checked (2026-11-15).
- Training data and knowledge cutoff: Not stated by Baidu on the pages checked (2026-11-15).
Teams that need documented risk evaluations should record this gap in their model assessment. See Model Cards and System Cards Explained.
Data handling
- Qianfan API training use and retention: Not stated by Baidu for ERNIE 5.1 on the pages checked (2026-11-15). Read Baidu AI Cloud's service agreement and privacy terms before sending personal or confidential data.
- Zero data retention: Not stated by Baidu on the pages checked (2026-11-15).
- Consumer chat: conversations in the ERNIE chat site follow Baidu's consumer terms, which are separate from Qianfan API terms.
- Jurisdiction: Qianfan is a Chinese cloud service, so China's generative AI measures and data rules apply to services offered in China; see the China Generative AI Measures and AI Labelling Rules Guide and China Cross-Border Data Transfers.
Qianfan API keys are non-human identities with spend and data exposure. See the LLMjacking Guide and How to set spend limits on LLM API keys.
Deprecation status
As of 15 November 2026, Baidu lists no deprecation notice, shutdown date or replacement for ERNIE 5.1 on the pages checked, and the Qianfan billing page updated on 11 November 2026 still prices it. Baidu does not publish a dedicated model deprecation schedule on the pages checked, so watch the Qianfan billing page and the ERNIE site for changes.
| Model or item | Details | Source |
|---|---|---|
| ERNIE-5.1 (Qianfan) | No notice date, shutdown date or replacement listed (2026-11-15); still priced | Qianfan billing page |
| ERNIE 5.1 (chat and AI Studio) | No retirement notice on the ERNIE site (2026-11-15) | ERNIE site |
See the LLM Deprecation and Retirement Tracker.
How ERNIE 5.1 differs from other models in its family
As of 2026-11-15. Sources: Baidu press release, ERNIE site and Qianfan billing page. Models listed by release date.
| Model | Released | Input / output per 1K (lowest tier) | Open weights | Positioning (Baidu wording) |
|---|---|---|---|---|
| ERNIE 5.0 | 2025-11-13 (preview) | See Qianfan billing page | No | Natively omni-modal model |
| ERNIE 5.1 | 2026-05-09 | CNY 0.004 / 0.018 | No | A Model That Writes Better and Understands You More |
The documented differences from ERNIE 5.0: ERNIE 5.1 is a compressed sub-network with about one third of the total and one half of the active parameters, lower inference cost, and a new post-training pipeline centred on on-policy distillation. Baidu describes ERNIE 5.0 as omni-modal and does not state ERNIE 5.1's modalities. The later ERNIE 5.5 Flash Preview and its open Base checkpoint are covered in Baidu ERNIE Models.
Practitioner checklist
- Record the Qianfan version name ERNIE-5.1 and the date adopted, and retest when Baidu updates the model or billing page.
- Confirm the context window and maximum output in the Qianfan console before designing long-context work; Baidu publishes pricing only up to 128K input.
- Model costs in CNY per 1,000 tokens with the 32K tier boundary included, and set spend alerts on Qianfan keys.
- Treat leaderboard ranks as dated provider statements and run your own evaluation on representative tasks.
- Read Baidu AI Cloud's service agreement and privacy terms, and decide which data classes may be sent to Qianfan.
- Assess China data and AI rules with counsel where your use falls under them.
- Store Qianfan API keys in a secrets manager, separate them by environment and rotate them on staff changes.
- Watch the Qianfan billing page and ERNIE site monthly for price or status changes.
Standards and references
- Baidu: ERNIE 5.1 Officially Released (accessed 2026-11-15)
- Baidu: ERNIE-5.1-Preview on LMArena (accessed 2026-11-15)
- Baidu AI Cloud: Qianfan billing and pricing (accessed 2026-11-15)
- Baidu: ERNIE site (accessed 2026-11-15)
- Baidu press release: ERNIE 5.0 unveiled at Baidu World 2025 (accessed 2026-11-15)
Related NHI Mgmt Group resources: Baidu ERNIE Models, Frontier AI Models Compared, LLM Release Tracker and Foundation Model Selection Guide.