AI BriefingPick
Baidu Releases ERNIE 5.5 Flash Base with 280B Parameters and 1M Context Support
·2026.11.05 09:00
Key point
The model features a Sparse MoE architecture with 14B active parameters per token and outperforms DeepSeek-V4-Flash-Base on MMLU and coding benchmarks.
Details
- Baidu released ERNIE-5.5-Flash-Base on November 5, 2026, as a pre-trained checkpoint for domain adaptation and RL research, supporting 1M token context during inference.
- The model utilizes a Sparse MoE structure with 280B total parameters and 14B active parameters per token, trained on 33.5T tokens from scratch.
- It achieves 93.3 on MMLU (5-shot) and 82.3 on HumanEval+ (Pass@1), surpassing DeepSeek-V4-Flash-Base (89.4 and 69.5) and Hy3-preview-Base (87.5 and 78.0).
- Architecture includes Hybrid Sparse Attention with MLA and HCA layers, Multi-Router Latent MoE with 512 routed experts, and a 200K vocabulary BPE tokenizer.
- Weights are available on Hugging Face in FP8 and BF16 formats under the ERNIE 5.5 Flash Base Model License Agreement v1.0.
- Post-trained ERNIE 5.5 Flash models are expected to launch soon, with the base model serving as a baseline for community fine-tuning and agent applications.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Moonshot AI releases Kimi K3.1 Pro open-weight model with 2.8T parameters