ERNIE 5.1

Overview

ERNIE 5.1 (文心 5.1) is Baidu's closed-weight flagship large language model, released on May 9, 2026. Baidu's release post opens: "Today, we officially release ERNIE 5.1. While inheriting the pre-training foundation of ERNIE 5.0, it compresses total parameters to approximately one-third and active parameters to approximately one-half, achieving leading foundational performance at its model scale using only about 6% of the pre-training cost of comparable models."

The headline is efficiency, not scale. ERNIE 5.0, announced as a "2.4 Trillion-Parameter Unified Multimodal Foundation Model" with a "<3% Activation Rate," was the ceiling; ERNIE 5.1 is a smaller mixture-of-experts model carved out of the same pre-training run. Baidu calls the technique Multi-Dimensional Elastic Pre-Training, combining elastic depth, elastic width / expert capacity, and elastic sparsity — the last defined as: "Through a variable Top-k routing mechanism, the number of activated experts is flexibly adjusted. Activating fewer experts reduces inference cost and improves decoding efficiency, while activating more enhances model capability."

Baidu says nothing about how many parameters that leaves. Baidu has never published an absolute parameter count for ERNIE 5.1 — only the "approximately one-third / approximately one-half" ratio. Any source quoting a hard number is not citing Baidu; this page does not state one.

Distribution is the sharpest contrast with the rest of the Chinese frontier. ERNIE 5.1 has no published weights, no Hugging Face repository, no OpenRouter listing, and no Artificial Analysis coverage. It exists as an API on Baidu's Qianfan platform, a chat box at ernie.baidu.com, and a set of leaderboard entries. Baidu's May 13 press release states that "a dedicated cluster powered by Baidu's proprietary Kunlunxin AI chip has supported the training of a key model in the ERNIE 5.1 series" — it stops short of saying ERNIE 5.1 itself was trained on it.

A note on naming

• "ERNIE X2" does not exist. Baidu's reasoning line runs ERNIE X1 → ERNIE X1 Turbo → ERNIE X1.1. No Baidu announcement, press release, or Qianfan model-list entry mentions an X2.
• Nothing newer than ERNIE 5.1 shipped as of July 8, 2026. Baidu's ERNIE blog index lists ERNIE 5.1 (May 9, 2026) as the most recent post; the ones before it are ERNIE-5.1-Preview (Apr 30), ERNIE-Image (Apr 15), and ERNIE 5.0 (Feb 6).
• "ERNIE," "Wenxin," "文心," and "Yiyan / 文心一言" all refer to the same family. ERNIE is the model; Wenxin/文心 is the Chinese brand; Yiyan is the consumer chat product.
• The API model ID is lowercase-hyphenated: ernie-5.1.

Capabilities

• Elastic MoE routing: A variable Top-k router changes how many experts fire per token, so serving cost and capability trade off at inference time rather than being fixed at training time.
• Agentic tool use: Baidu reports that "on the τ³-bench and SpreadsheetBench-Verified agent evaluation tasks, ERNIE 5.1 surpasses DeepSeek-V4-Pro."
• Search-grounded answering: ERNIE 5.1's strongest independently visible result is on arena.ai's search leaderboard, where it ranks 5th — the only Chinese model in that top five.
• Tool-augmented mathematics: "On AIME26 (with tool use), a challenging mathematical competition benchmark, ERNIE 5.1 scores 99.6 — second only to Gemini 3.1 Pro."
• Long-form Chinese and English writing: Baidu's post is subtitled "A Model That Writes Better and Understands You More," and claims creative-writing quality approaching Gemini 3.1 Pro on internal evaluations.
• 128K context with 64K output: The ernie-5.1 entry in Baidu's Qianfan model list specifies a 128K context window and a max-output range of [1, 65536].

Technical Specifications

• API model ID: ernie-5.1
• Architecture: sparse Mixture-of-Experts transformer with variable Top-k expert routing
• Total parameters: not published. Baidu states only "approximately one-third" of ERNIE 5.0's 2.4T.
• Active parameters: not published. Baidu states only "approximately one-half" of ERNIE 5.0's (itself given only as "<3% activation rate").
• Context window: 128K tokens
• Max output: 1–65,536 tokens
• Modality: text generation (Qianfan classifies ernie-5.1 under 文本生成, not multimodal)
• Pre-training: "Multi-Dimensional Elastic Pre-Training" — elastic depth, elastic width / expert capacity, elastic sparsity
• Post-training: four stages — unified SFT, parallel domain-expert training, on-policy distillation, and general online reinforcement learning
• RL infrastructure: disaggregated, fully-asynchronous design with FP8 training/inference consistency and heterogeneous resource scheduling
• Training hardware: not disclosed for ERNIE 5.1 specifically. Baidu's May 13, 2026 press release says only that "a dedicated cluster powered by Baidu's proprietary Kunlunxin AI chip has supported the training of a key model in the ERNIE 5.1 series."
• Weights: not released. License: proprietary.

Baidu does not publish a knowledge cutoff, a training-token count, or a parameter count for ERNIE 5.1, so this page does not state one.

Use Cases

• Search-grounded assistants: The one arena.ai board where ERNIE 5.1 is genuinely near the top is search (rank 5), which matches Baidu's core business and its own framing of the model.
• Chinese-language enterprise deployment: Qianfan is the practical path for organizations already inside Baidu AI Cloud, and pricing is quoted in CNY for a CNY-billed account.
• Spreadsheet and workflow agents: Baidu specifically cites SpreadsheetBench-Verified and τ³-bench as places ERNIE 5.1 beats DeepSeek-V4-Pro — an agentic-workflow profile rather than a raw-reasoning one.
• Tool-using math and analysis: The AIME26 score is reported with tool use, so the relevant deployment shape is a model with a calculator/interpreter attached, not a bare chat completion.
• Long-document processing: 128K in, up to 64K out, in a single call.
• Consumer and creative products: Baidu says the model is "progressively rolled out on over ten creative production agent platforms," including ISEKAI ZERO and Mulan AI.

Performance / Benchmarks

Baidu's ERNIE 5.1 release post presents its benchmark comparison as an image, not as machine-readable text. Only these figures appear in the post's prose, and they are all vendor-reported:

Baidu publishes no numeric GPQA or MMLU-Pro score. This page does not state one.

arena.ai placements (independent, fetched July 8, 2026)

arena.ai — the former LMArena, rebranded in January 2026 — is the only third-party leaderboard carrying ERNIE 5.1. Its text board, in a snapshot dated July 1, 2026, covers 369 models and 7,152,929 votes:

On the same text board, the Chinese models above ERNIE 5.1 are qwen3.7-max-preview (15, 1475), glm-5.1 (22, 1472), and glm-5.2 (max) (26, 1469). Below it: mimo-v2.5-pro (29), kimi-k2.6 (34), deepseek-v4-pro (38), minimax-m3 (54).

This partly contradicts a live Baidu claim. Baidu's May 13, 2026 press release states ERNIE 5.1 "ranked #1 among Chinese models on LMArena's text and search leaderboards," and the April 30 preview post claimed "No. 1 among Chinese models and No. 13 globally." On the text board, neither holds today: as of the July 1 snapshot, ERNIE 5.1 is the fourth-highest-scoring Chinese model and 27th overall. On the search board the "#1 among Chinese models" half of the claim does still hold — no Chinese model outranks it — though the release post's "4th place globally" has slipped to 5th as vote counts grew. Preliminary rankings move; treat launch-day placements as snapshots, not standings.

Artificial Analysis does not track any Baidu or ERNIE model, so there is no third-party intelligence-index or independent latency/throughput measurement for ERNIE 5.1. (Artificial Analysis and arena.ai use different scales and are never comparable.)

Limitations

• Closed weights, full stop. No Hugging Face repository, no OpenRouter listing, no self-hosting, no fine-tuning on your own hardware. DeepSeek, Qwen, GLM, Kimi, MiniMax and Tencent all ship downloadable checkpoints for their current flagships; Baidu does not.
• Baidu open-sources only previous-generation language models. baidu/ERNIE-4.5-300B-A47B-PT and the rest of the ERNIE 4.5 family are Apache 2.0 on Hugging Face — first published June 2025, with an ERNIE-4.5-21B-A3B-Thinking variant added that September. ERNIE 5.0 and 5.1 weights have never been published. (Outside the LLM line Baidu does ship current-generation weights: baidu/ERNIE-Image, Apache 2.0, April 2026.)
• No published parameter count. You cannot size a deployment, estimate cost-per-FLOP, or compare efficiency claims against DeepSeek or Qwen, because Baidu gives only a ratio against another model whose active-parameter count is itself given only as "<3%."
• Every benchmark number is vendor-reported, and the comparison table is a bitmap. GPQA and MMLU-Pro get an adjective ("approaches the performance of leading closed-source models"), not a score. The 6% pre-training-cost claim has no independent replication and no disclosed accounting method.
• The vendor's own text-leaderboard claim is stale. "#1 among Chinese models on LMArena's text leaderboard" was true for the preview in late April; on the July 1, 2026 arena.ai snapshot the model is 27th overall and behind three other Chinese entries. (The parallel claim on the search board does still hold.)
• No independent third-party evaluation infrastructure. Artificial Analysis does not cover it; OpenRouter does not serve it. There is no neutral latency, throughput, or price-performance measurement to check Baidu's efficiency story against.
• Pricing is CNY-only, and access is China-centric. Qianfan bills in 元. There is no USD rate card, and the practical onboarding path assumes a Baidu AI Cloud account.
• On the Hugging Face Hub, Baidu reads as an OCR company, not a frontier lab. Its most-downloaded asset by a wide margin is baidu/Unlimited-OCR (MIT, ~1.08M downloads in 30 days) — a 3B document-parsing VLM — followed by baidu/Qianfan-OCR (~256K). The largest ERNIE 4.5 flagship, ERNIE-4.5-300B-A47B-PT, drew about 2,100.

Pricing & Access

There is no free tier of weights. ERNIE 5.1 cannot be self-hosted.

Hosted access is sold through Baidu's Qianfan platform. Baidu's 模型服务计费 ("Model Service Billing") page, last updated 2026-07-07, lists ERNIE 5.1 under 推理服务 ("inference service") with prompt-length-tiered pricing:

All figures are yuan (CNY). Baidu publishes no USD rate card for ERNIE 5.1, and no conversion is implied here. For reference, the same page prices ERNIE 5.0 higher at every tier (0.006 / 0.024 元/千 tokens at ≤32K), which is consistent with 5.1 being the smaller model.

Consumer access: the chat interface at ernie.baidu.com, and the Baidu AI Studio playground. Baidu says the model is "progressively rolled out on over ten creative production agent platforms."

Other ERNIE 5.x model IDs on Qianfan: ernie-5.0, ernie-5.0-thinking-preview, ernie-5.0-thinking-latest. All share the same 128K context and [1, 65536] output range.

Ecosystem & Tools

• ERNIE 5.1 release post — Baidu's primary announcement, May 9, 2026
• Qianfan 模型列表 — the authoritative ernie-5.1 spec: context, max output, category
• Qianfan 模型服务计费 — the CNY rate card
• ernie.baidu.com — consumer chat, free access to the latest ERNIE model
• Baidu on Hugging Face — ERNIE 4.5 (Apache 2.0), ERNIE-Image, Unlimited-OCR. No ERNIE 5.x.
• baidu/ERNIE-Image — Apache 2.0, 8B single-stream DiT text-to-image model, the current-generation open release
• Kunlunxin (昆仑芯) — Baidu's in-house AI accelerator; a Kunlunxin cluster "supported the training of a key model in the ERNIE 5.1 series," per Baidu's May 13 press release

Community & Resources

• ERNIE 5.1 Officially Released! — Baidu's official post, May 9, 2026
• ERNIE-5.1-Preview Tops LMArena Text Leaderboard as No.1 Chinese Model! — the April 30 preview claim, now superseded
• ERNIE 5.0: A 2.4 Trillion-Parameter Unified Multimodal Foundation Model — the model ERNIE 5.1 was compressed from
• Baidu Advances Agent Portfolio… — Create 2026 press release, May 13, 2026 (Kunlunxin training cluster, DAA metric)
• Baidu releases Ernie 5.1 with pre-training cost at only 6% of industry peers — CnTechPost
• 百度发布文心 5.1：搜索能力登顶国内，预训练成本仅为业界 6% — 量子位 (QbitAI)
• arena.ai text leaderboard and search leaderboard
• Compare with DeepSeek V4, Qwen3.7-Max, GLM-5.2, Kimi K2.6, Tencent Hy3, and MiniMax-M3