AI Model Timeline

Tracking the accelerating release frequency of frontier AI models.

ERNIE 5.5 Flash Base

Baidu·

LLMsopen-weightMulti-GPU Self-Host

Overview

Open pretrained base checkpoint of ERNIE 5.5 Flash: a 280B total / 14B active sparse MoE (43 layers, 512 routed experts with top-10 routing plus one shared expert), text-only, pre-trained from scratch on 33.5T tokens. This is the end-of-pre-training checkpoint with no instruction tuning or post-training; it is served through the completions API, is not chat-ready, and Baidu positions it as a starting point for domain adaptation, instruction tuning and RL research. The post-trained model is the separate, API-only ERNIE 5.5 Flash Preview. Context 131K in pre-training, 1M enabled in the released config. FP8 (~288 GB) and BF16 (~573 GB) safetensors; Baidu states no VRAM requirement. ERNIE 5.5 Flash Base Model License Agreement v1.0: the grant does not cover the European Union, the United Kingdom or South Korea, commercial use needs a notice to Baidu, and above US$50M related annual revenue prior written authorization. Baidu's figures are few-shot base-model evaluations (MMLU 93.3 and GPQA 60.2 at 5-shot, among others) that are not comparable with the chat-model values tracked here, so no benchmark field is filled.

Capabilities and innovations

280B total / 14B active parameters (sparse MoE)Pretrained base model, completions API onlyText-only, up to 1M-token contextFP8 and BF16 safetensorsERNIE 5.5 Flash Base Model License (excludes EU/UK/South Korea)Open release of a frontier-scale pre-training checkpoint before the post-trained modelPre-trained from scratch on 33.5T tokens

Benchmarks

No benchmark values are recorded for this model.

  • GPQA Diamond: not reported by the vendor — Baidu prints 'GPQA' 60.2 as a 5-shot exact-match base-model evaluation, without 'Diamond'. A few-shot completion score is not comparable with the GPQA Diamond values tracked here and is not entered.
  • Humanity’s Last Exam (no tools): not reported by the vendor
  • SWE-bench Verified: not reported by the vendor
  • SWE-bench Pro: not reported by the vendor
  • MMLU: not reported by the vendor — Baidu prints MMLU 93.3 and MMLU-Redux 94.1 as 5-shot base-model evaluations from its own harness, compared only against other base checkpoints; not entered.
  • MMLU-Pro: not reported by the vendor — Baidu prints MMLU-Pro 78.4 as a 5-shot base-model evaluation; not entered, for the same reason as MMLU.
  • Terminal-Bench 2.x: not reported by the vendor
  • MMMU-Pro: not reported by the vendor — Not applicable — the model is text-only.

Architecture and hardware

Parameters
280B total, 14B active per token (MoE)
Estimated VRAM at Q4
~161 GB, Multi-GPU Self-Host class
Quantization formats
FP8, BF16
Recommended runtime
SGLang
License
ERNIE 5.5 Flash Base Model License v1.0

Links

More from Baidu

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.