---
title: "Baidu Cuts ERNIE's Pre-training Bill to a Fraction, Then Rebuilds Its Reinforcement Learning"
date: 2026-11-17
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-ernie51-training-efficiency
language: en
---

# Baidu Cuts ERNIE's Pre-training Bill to a Fraction, Then Rebuilds Its Reinforcement Learning

> ERNIE 5.1 inherits a larger model's knowledge at a fraction of the training cost, then retrains how agents are taught.

A model update usually arrives with a bigger number attached. Baidu's latest release arrives with a smaller one. ERNIE 5.1, launched on 9 May 2026, is described by the company as costing about 6% of the pre-training compute of a comparable model of its size — not because it is weak, but because it was extracted from a stronger parent rather than trained from scratch.

The claim is Baidu's own, and it comes with an architectural story attached. Together they describe a shift in how Chinese labs think about the large model (大模型) business: the scarce resource is no longer just capability, but the compute needed to buy it.

## Taking the best parts of a bigger model

ERNIE 5.1 is built on ERNIE 5.0, Baidu's 2.4-trillion-parameter native full-modal model released in final form in January 2026. Instead of training a smaller model independently, Baidu's team used what it calls a Once-for-All elastic training framework. During a single pre-training run of the parent, the method dynamically samples many submodels of different sizes, so that one training pass produces a matrix of candidates rather than one fixed network.

Three axes of elasticity are involved:

- **Depth.** The trainer randomly skips Transformer layers, so submodels of different depths share weights.

- **Width.** The pool of experts actually used in mixture-of-experts (混合专家, MoE) layers is randomly masked, forcing the remaining experts to take on more varied work.

- **Sparsity.** A variable top-k routing mechanism changes how many experts activate per token, trading inference cost against capability.

From that matrix, Baidu extracted a submodel with roughly one-third of ERNIE 5.0's total parameters and about half its activated parameters. The company says the result retains the parent's knowledge and uses about 6% of the pre-training compute of peers at the same scale. That percentage is a self-reported comparison, not an externally audited figure, and the "comparable model" baseline is not named in the release post.

## The reinforcement learning plumbing

The more substantive engineering is in post-training. Baidu rebuilt the reinforcement learning (强化学习) stack around what it calls a disaggregated fully-asynchronous architecture, constructed on its PaddlePaddle framework. The control planes of four subsystems — training, inference, reward and agent loop — are decoupled, connected through high-performance network components. Each can be deployed and scaled independently, and the pipeline stages can overlap.

Two details are worth noting. A unified FP8 operator library narrows the precision gap between training and inference, and an optimized technique called Rollout Router Replay (R3) reduced KL divergence by 50%, according to Baidu, without adding meaningful inference time. A pooled-CPU strategy puts otherwise idle cluster CPUs to work on code sandboxes and verifiers.

This is infrastructure built for agentic training — long-horizon tasks where a model acts, checks results and acts again. Baidu argues that serial post-training pipelines have become a bottleneck, and that trying to fuse every capability in one stage produces a see-saw effect in which one skill improves at another's expense.

## A four-stage recipe

The post-training pipeline is described as four stages:

- **Unified supervised fine-tuning (SFT)** on high-quality, multi-domain instruction data, establishing instruction-following and tool use.

- **Parallel domain-expert training** for code, reasoning and agents, each with its own reward signal and algorithm, so the tasks do not interfere.

- **On-policy distillation (OPD).** A single student model samples from its own distribution and learns from several expert teachers through token-level reverse KL, folding their abilities into one parameter space.

- **General online reinforcement learning** for open-ended chat and creative writing, where Baidu says distillation tends to smooth the output distribution and is replaced with RL.

The pattern — train specialists in parallel, then merge them — is a practical answer to a real problem. It lets a lab iterate on an agent or a coding model without retraining everything, which matters when the capabilities keep multiplying.

## What the leaderboards say

Baidu reports that on 9 May 2026, ERNIE 5.1 scored 1,223 on the Arena search leaderboard, placing fourth globally and first in China. It also reports 99.6 on AIME26 with tools, behind only Gemini 3.1 Pro, and strong results on τ³-bench, SpreadsheetBench-Verified, GPQA and MMLU-Pro.

These are vendor-reported results on named benchmarks, which makes them more checkable than a vague claim of "leading performance" — but they are still the vendor's numbers. The model is available through Baidu's Yiyan site and its Qianfan cloud platform, and on Baidu's AI Studio community, so independent testing is at least possible.

## Honest limitations

- The 6% pre-training-cost figure and the 50% KL reduction are Baidu's own measurements, with no published independent audit or a fully specified comparison baseline in the release post.

- Leaderboard scores, including the Arena 1,223 figure, are a snapshot. Arenas are crowdsourced and rankings move as new models are submitted.

- The release compares ERNIE 5.1 against model versions and closed systems that are themselves moving targets; those comparisons date quickly.

- No pricing for ERNIE 5.1 was stated in the source reviewed for this article, so the economic claim cannot be checked against an API bill.

## What readers can do now

- **Read "trained from scratch" as a cost claim, not a quality claim.** Ask whether a smaller model was distilled or extracted from a larger parent before comparing compute budgets.

- **Watch the post-training pipeline, not just the final score.** Parallel expert training plus distillation is becoming a common recipe; it tells you how fast a vendor can add a new skill.

- **Verify leaderboard positions yourself.** Check the current Arena tables rather than quoting a launch-day rank in a procurement document.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-ernie51-training-efficiency
Free to quote with attribution and a link to the original.
