ERNIE 5.5 Flash Base: Baidu's 280B open checkpoint for further training
Baidu released the weights of ERNIE-5.5-Flash-Base, a 280B-total / 14B-active sparse MoE base model pre-trained on 33.5T tokens, as a starting point for domain fine-tuning, instruction tuning and RL research.
Why it matters
Baidu's first open hundred-billion-parameter model in 16 months: a raw pre-training checkpoint with no chat tuning that tops DeepSeek-V4-Flash-Base on 24 of 30 benchmarks in Baidu's own unified evaluation, with a 1M-token inference context.
What you could build with it
A research team studying post-training methods could take this fixed pre-training checkpoint and run controlled comparisons of instruction-tuning recipes against DeepSeek-V4-Flash-Base, since both start from the same stage and Baidu publishes a unified evaluation harness.
Does it hold up?
Too early to judge: weights landed today, serving is documented for SGLang and vLLM, but no independent benchmarks or community fine-tunes exist yet, and the custom licence bars use in the EU, UK and South Korea.
Built with ERNIE 5.5 Flash Base
- Baidu Releases ERNIE 5.5 Flash Base for Further Trainingarticle · Unite.AI covers the architecture (43 layers, hybrid HCA/MLA attention, 512 routed experts), benchmarks and the restrictive custom licence.
- Baidu Drops ERNIE 5.5 Flash, a raw 280B open checkpoint beating DeepSeekarticle · AlphaSignal notes the model has no chat alignment built in, the FP8 weights alone need roughly 280GB, and Baidu's scores are not independently reproduced.
- Baidu open-sources the ERNIE 5.5 Flash model with 280 billion parametersarticle · ChainCatcher report on the release: first open Baidu model of this scale since the ERNIE 4.5 series in June 2025, with the stricter custom licence.
- baidu/vLLM-ERNIEgithub · 39 ★ · Baidu's vLLM plugin for serving ERNIE 5.5 Flash Base, referenced in the official model card's deployment guide.
- PaddlePaddle/FlashMaskgithub · 58 ★ · The FlashMask-4 attention kernel library behind ERNIE's training stack: CuTe-DSL kernels for Hopper and Blackwell, up to 175% faster forward than FlashAttention-4 MaskMod on SM100.
Learn more
First spotted on article: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MNVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4…models · JEV 0.72Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72NVIDIA Cosmos 3.5 world models: now openly availableNVIDIA's Cosmos 3.5 family of open world foundation models for robots, autonomous vehicles and vision agents is now…models · JEV 0.71
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.