Ernie 5.1: Smaller, More Efficient, Stronger

ERNIE 5.1 is not just an incremental update. Baidu decided to pursue a significant increase in efficiency while maintaining — or even improving — performance. The model is built on a Mixture of Experts (MoE) architecture, where each query activates only a portion of the parameters. The total number of parameters is approximately one-third of ERNIE 5.0, with active parameters during each computation being about half.

A crucial economic figure: training ERNIE 5.1 cost only 6% of the expenses of comparable frontier models. In other words, Baidu reduced pre-training costs by 94%. This is a game-changing number — especially in the context of US sanctions on chip exports to China, which force Chinese companies to work with more limited computational resources.

Two key technical innovations stand behind this result. Baidu created a so-called Once-For-All elastic training framework — a methodology that allows extracting an optimal sub-network from the broader ERNIE 5.0 architecture without needing to train the model from scratch. This is complemented by a new infrastructure for disaggregated, fully asynchronous reinforcement learning, which further streamlines the post-training phase focused on agentic capabilities.

Benchmark Results: Where ERNIE 5.1 Outperforms DeepSeek

The numbers are telling. In the popular evaluation environment LMArena, ERNIE 5.1 Preview achieved a score of 1,476 points in the Text Arena category, ranking 13th globally — and 1st among all Chinese models. In the Arena Search category, it scored 1,223, taking 4th place worldwide.

A specific comparison with DeepSeek-V4-Pro shows where ERNIE 5.1 truly leads:

• τ³-bench (tau3-bench) — agentic benchmark simulating complex autonomous tasks: ERNIE 5.1 outperforms DeepSeek-V4-Pro
• SpreadsheetBench-Verified — test of working with tabular data in an agentic environment: ERNIE 5.1 leads again
• Legal & Government category on LMArena: ERNIE 5.1 took 1st place globally, ahead of all foreign models
• AIME26 (mathematical competition tasks with tool assistance): score of 99.6 — second best in the world, only behind Gemini 3.1 Pro
• Math category on LMArena: 9th place globally

On knowledge and reasoning benchmarks GPQA and MMLU-Pro, ERNIE 5.1 approaches leading closed models — specifically Gemini 3.1 Pro. The same applies to creative writing in internal evaluations.

Where Does ERNIE 5.1 Still Lag?

For an objective picture, it's necessary to state where ERNIE 5.1 falls short of the world's top models. Its overall ranking on LMArena places it behind Gemini 3.1 Pro, Claude Opus 4, and leading OpenAI models. In areas like multimodal image processing or video models, it does not yet stand behind ERNIE 5.0 — ERNIE 5.1 is a purely text-based model.

DeepSeek, meanwhile, still dominates in the area of open-source models with Chinese localization and in the efficiency of inference costs for the developer community. ERNIE 5.1 is primarily a closed commercial model available through Baidu platforms.

Access and Price: What Developers Need to Know

ERNIE 5.1 is available in three ways:

• ERNIE Chat UI — web interface for regular users at ernie.baidu.com
• Baidu AI Studio ERNIE 5.1 Playground — testing environment for developers
• Qianfan API — for enterprise and developer integrations

The price via Qianfan API is set at $0.59 per million input tokens and $2.65 per million output tokens. The context window is 128,000 tokens. Compared to similarly performing models (Claude Opus 4, Gemini 3.1 Pro), this is a competitive price, especially for companies with a large volume of queries.

Availability for Czech and European Companies

Here, it's necessary to be realistic. ERNIE 5.1 is hosted on Baidu's Chinese infrastructure. For European companies that must comply with GDPR requirements and especially strict interpretations of EU cross-border data transfer regulations, using the API via Qianfan can be problematic. Data processed in Chinese data centers is not subject to the European legal framework.

The ERNIE interface is not available in Czech, and the model is primarily optimized for Chinese and English. For Czech users, it is therefore more of an interesting benchmark player than a practical everyday tool — at least for now.

Why ERNIE 5.1 is Important for the Entire AI Industry

The main message of ERNIE 5.1 is not its absolute performance — it is a proof of concept for efficient training. If top-tier performance can be achieved at 6% of standard costs, the economics of the entire industry change. We observe this trend also with DeepSeek V3/V4, Mistral models, and Meta with Llama 4 — both Chinese and European research shows that "bigger = better" is no longer absolutely true.

For Baidu, ERNIE 5.1 also has a strategic dimension: the company faces fierce domestic competition from Alibaba (Qwen), ByteDance, and DeepSeek itself. Proving that it can train a world-class model at a fraction of the cost, even with restrictions on access to NVIDIA chips, is crucial for it in terms of investors and regulators.

According to Baidu's official blog, further updates to ERNIE models are expected at the Baidu Create 2026 conference — so ERNIE 5.1 is not the final stop, but the starting gun for a new generation.