Baidu Releases ERNIE 5.1 Featuring High Parameter Efficiency and Asynchronous RL Architecture

ArticleBaidu AI TeamAnnouncement

Baidu has officially released ERNIE 5.1, compressing total parameters to roughly one-third and active parameters to approximately one-half compared to ERNIE 5.0, while requiring around 6% of the pre-training compute cost of comparable models.

Key points

  1. Efficient pre-training: Derived from ERNIE 5.0 via a Once-For-All elastic training framework that optimizes elastic depth, width, and MoE sparsity, cutting pre-training compute to 6% of comparable models.
  2. Disaggregated asynchronous RL infrastructure: Built on PaddlePaddle, decoupling the control plane into four independently scalable subsystems: training, inference, reward, and agent loop.
  3. Training-inference consistency: Unified FP8 low-precision operators and optimized Rollout Router Replay (R3) reduce MoE routing divergences.
  4. Four-stage post-training pipeline: Incorporates unified SFT, parallel domain expert training, Multi-Teacher On-Policy Distillation (MOPD), and general online RL for high-entropy creative and conversational tasks.
  5. Benchmark metrics: Scored 1,223 to secure 4th place globally on the Arena Search leaderboard (1st among Chinese models), outperformed DeepSeek-V4-Pro on agent benchmarks (τ³-bench, SpreadsheetBench-Verified), and scored 99.6 on AIME26.

Why it matters

As models scale, pre-training and multi-task reinforcement learning face severe cost bottlenecks and conflicting optimization goals. Baidu's approach illustrates how elastic architectural extraction and decoupled RL pipelines can substantially reduce computational overhead while maintaining high reasoning and agent performance.

Analysis

Baidu targets key technical constraints in large MoE and agent training. Decoupling the RL architecture enables independent resource allocation for inference, training, and rewards, supported by elastic CPU pooling for sandboxed verification. The four-stage post-training framework solves the classic 'seesaw' effect of sequential fine-tuning: expert models are trained independently and merged via on-policy distillation, leaving open-ended creative tasks to a subsequent online RL phase that avoids excessive probability distribution smoothing.

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:xAI

    xAI Releases Grok 4.1

    xAI has launched Grok 4.1 across grok.com, X, and mobile applications. The update focuses on real-world conversational usability, emotional intelligence, and reduced hallucination rates.

    AI & LLMs· Announcement

  • Link:primeintellect.ai

    Mixedbread Trains Toast 1 Deep-Search Agent Using Prime-RL

    A case study by Prime Intellect outlines how Mixedbread trained Toast 1 using Prime-RL and Verifiers v1. Fine-tuned from Qwen3.6-35B-A3B, the retrieval subagent matches frontier-level context gathering 12 times faster and at significantly lower cost.

    AI & LLMs· Resource

  • Link:Moonshot AI

    Moonshot AI Releases Kimi K3: Open-Weight 2.8T MoE Model

    Moonshot AI has released Kimi K3, an open-weight native multimodal agentic model featuring 2.8 trillion total parameters and a 1-million-token context window. The architecture uses a sparse MoE setup that activates 104 billion parameters per token.

    AI & LLMs· Announcement

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.