On November 5, 2026, Baidu's ERNIE team released the open weights of ERNIE 5.5 Flash Base — a 280-billion-parameter mixture-of-experts model with only 14B active per token. It is a pre-training checkpoint, not a finished chat model: a foundation others fine-tune, not an app you can use today.
2026-11-05 · by Moe Ameen
On November 5, 2026, Baidu's ERNIE team open-sourced ERNIE 5.5 Flash Base, publishing the model's weights on Hugging Face and ModelScope. The important word is "Base": Baidu released the checkpoint at the end of pre-training, before any mid- or post-training. Its tagline — "an efficient starting point for what comes next" — states the intent plainly. This is a foundation for builders and researchers to fine-tune and post-train, not an instruction-tuned assistant you can chat with out of the box.
Architecturally it is a sparse mixture-of-experts model: 280 billion total parameters but only about 14 billion activated per token, which is what keeps inference cheap and fast relative to its size. Baidu lists 43 Transformer layers with a hidden size of 4096, a roughly 200K-entry tokenizer, pre-training from scratch on 33.5 trillion tokens, a training context extended to 128K, and support for a 1M-token context in fine-tuning and inference. Weights ship in FP8 and BF16 safetensors formats under the "ERNIE 5.5 Flash Base Model License Agreement v1.0" (read the license terms before deploying), and the model can be served through Hugging Face Transformers, SGLang, and vLLM, with a fine-tuning example in NVIDIA's Megatron-Bridge.
Baidu's own release benchmarks it against other recent base checkpoints — DeepSeek-V4-Flash-Base, a Tencent Hy3-preview base, and Xiaomi's MiMo-V2.5-Base — under a shared harness, and reports leading scores on most of about 30 tests (for example MMLU 93.3, HumanEval+ 82.3, and RepoQA 91.7), while trailing on a few Chinese-QA and long-context benchmarks. Treat these as the vendor's self-reported figures pending independent reproduction. Two practical caveats matter for anyone tempted to run it: it is a raw base checkpoint with no instruction tuning, and the FP8 weights alone are roughly 288GB, so this is a data-center model, not something you load on a laptop. The release lands in a broader 2026 race among Chinese labs to open-weight efficient, low-cost models.
The honest thing to say about ERNIE 5.5 Flash Base is that there is no "use it today" for a creator — it is a pre-training checkpoint, so until a lab or vendor post-trains it into an instruction-following model and wraps it in an API, it produces unpredictable raw text and nothing else. The news to act on is the trend it represents, not the file you can download: cheaper, faster inference keeps commoditizing the drafting half of content, which means the part that actually decides whether you ship — turning words into finished, on-brand, scheduled posts — is where the leverage sits. [Kompozy](/) is built for that half, and it already runs on frontier models (Claude and OpenAI for copy, gpt-image and Gemini for images, HeyGen for avatar video), so you do not have to wait for anyone to tame a base checkpoint to get the benefit.
Concretely: hand Kompozy an idea or a rough draft and it fans it into formats a language model cannot render — a captioned [Persona Short](/glossary/persona-shorts), a brand-exact [Carousel](/glossary/hyperframes), a [Quote Graphic](/glossary/persona-tweet) from the strongest line, a Blog Article and an Email Newsletter — all held to one [Persona Brief](/glossary/persona-brief) voice, then [Autopilot](/glossary/autopilot) reframes each to 9:16, 1:1, and 16:9 and publishes across the eight social platforms plus blog and email behind a per-post review gate. When a fast, cheap ERNIE-derived API eventually shows up, it slots in upstream as one more drafting option; the engine that finishes and distributes the work stays the same.
ERNIE 5.5 Flash Base is an open-weight base language model from Baidu's ERNIE team, released November 5, 2026. It is a sparse mixture-of-experts model with 280B total parameters and about 14B activated per token, pre-trained from scratch on 33.5 trillion tokens. Crucially it is a pre-training checkpoint — released before instruction tuning — so it is a foundation to fine-tune, not a finished chat assistant.
Not directly. As a base checkpoint it has no instruction tuning, so it will not reliably follow a prompt until someone post-trains it, and running the weights at all needs data-center GPUs. To make and publish content today, use a product built on already-post-trained frontier models — Kompozy, for example, generates captioned video, carousels, images, blogs, and newsletters and publishes them across platforms.
The weights are published openly on Hugging Face and ModelScope (in FP8 and BF16 formats) under the ERNIE 5.5 Flash Base Model License Agreement v1.0. Read that license before any production or commercial deployment, as model licenses often carry use and regional conditions.
Baidu's release compares it against other recent base checkpoints — including DeepSeek-V4-Flash-Base, a Tencent Hy3-preview base, and Xiaomi's MiMo-V2.5-Base — under a shared harness, reporting leading scores on most of about 30 benchmarks while trailing on some Chinese-QA and long-context tests. Those are Baidu's own numbers; independent reproduction had not been published at release.