ERNIE 5.1

1. Agent Performance

This was the area where I noticed the biggest improvement.

Single-turn chat does not always reveal much about a model. Multi-step tasks do. Ask a model to gather information, run calculations, use tools, organize the results, and produce a table, and weaknesses start showing up quickly.

ERNIE 5.1 handled these workflows more cleanly. Information from earlier steps carried through to later ones, and longer task chains were less likely to fall apart halfway through.

On τ³-bench and SpreadsheetBench-Verified, ERNIE 5.1 outperformed DeepSeek-V4-Pro.

That matches my experience with it. As an agent model, it feels less like something that simply answers prompts and more like something that can stay with a task until it is finished.

2. Reasoning and Math

On the AIME26 math benchmark with tool use enabled, ERNIE 5.1 scored 99.6, second only to Gemini 3.1 Pro.

It also performed close to leading closed models on benchmarks such as GPQA and MMLU-Pro.

Benchmarks are not the same as real-world performance, but these results do show that complex reasoning is not a weak spot.

For coding logic, multi-step calculations, and longer chains of reasoning, ERNIE 5.1 no longer needs to be judged on a softer “good for a domestic Chinese model” curve. It is competitive with top-tier models on its own terms.

3. Knowledge and Creative Writing

ERNIE 5.1 also scores well on knowledge-heavy benchmarks including MMLU-Pro and GPQA. Baidu’s internal testing places its creative writing performance close to Gemini 3.1 Pro.

I tried it on copywriting, rewriting, and expansion tasks. The clearest improvement was not more elaborate language. The writing simply felt smoother.

ERNIE 5.0 could sometimes fall into familiar model habits: forced transitions, filler phrases, or paragraphs that always started with a summary before getting to the point. ERNIE 5.1 still has traces of that, but I ran into it less often, and the output needed less cleanup afterward.

The 128K context window is enough for normal long-form documents. For very large knowledge bases or documents running into the million-character range, you will still need to split the material.

4. What Changed From ERNIE 5.0

The most interesting change is that ERNIE 5.1 is actually smaller.

Its total parameter count has been reduced to one-third of ERNIE 5.0, while active parameters are down to half. Baidu says its pretraining cost is only 6% of comparable models.

The model uses an Once-For-All elastic training framework. During training, multiple subnetworks of different sizes are optimized together, allowing Baidu to extract a smaller and more efficient architecture afterward.

The result is straightforward: the model is smaller, while some capabilities are stronger than ERNIE 5.0.

That matters more in real products than another jump in raw parameter count. Agent systems and automated workflows may call a model many times for a single user request. A small difference in cost per call becomes much more important once usage scales.

ERNIE 5.1 has a fairly clear set of strengths. Its agent performance is strong, its reasoning is competitive with leading models, its Chinese writing has improved, and its API pricing makes it attractive for developers, small and midsize businesses, and workloads that involve a lot of model calls.

It is not the right fit for everything. A 128K context window can be restrictive for very large document sets, output is capped at 16K tokens, and native multimodal capability is not its main strength. If your work regularly depends on 1M-token context windows or heavy image and video understanding, other models will make more sense.

For agent workflows, complex reasoning, automation, or cost-sensitive API usage, ERNIE 5.1 is worth testing with real tasks.

I would judge it on three numbers: task completion rate, average token usage, and cost per completed task.

Those will tell you more than a benchmark leaderboard.