Baidu's open-weight sparse-MoE base model — 280B total parameters with ~14B active per token — built for cheap, fast inference and released as a pre-training checkpoint for builders to fine-tune.
Last verified · 2026-11-05 · by Moe Ameen
ERNIE 5.5 Flash Base is an open-weight base language model from Baidu's ERNIE team, open-sourced on November 5, 2026. The "Base" matters: Baidu released the checkpoint at the end of pre-training, before any mid- or post-training, and describes it as "an efficient starting point for what comes next." In other words it is a foundation other teams fine-tune and post-train — not an instruction-tuned assistant you can prompt out of the box the way you would ChatGPT or ERNIE's own chat products.
Under the hood it is a sparse mixture-of-experts model: 280 billion total parameters but only about 14 billion activated per token, which is the design choice behind the "Flash" name — it keeps inference cheap and fast for a model of its size. Baidu lists 43 Transformer layers, a hidden size of 4096, a roughly 200K-entry tokenizer, pre-training from scratch on 33.5 trillion tokens, a 128K training context, and support for a 1M-token context in fine-tuning and inference. Weights are published on Hugging Face and ModelScope in FP8 and BF16 safetensors under the ERNIE 5.5 Flash Base Model License Agreement v1.0, and the model runs through Hugging Face Transformers, SGLang, and vLLM, with a fine-tuning example in NVIDIA's Megatron-Bridge.
Baidu's release benchmarks it against other base checkpoints — DeepSeek-V4-Flash-Base, a Tencent Hy3-preview base, and Xiaomi's MiMo-V2.5-Base — reporting leading scores on most of about 30 tests (for example MMLU 93.3, HumanEval+ 82.3, RepoQA 91.7) while trailing on a few Chinese-QA and long-context benchmarks. Those are vendor-reported numbers pending independent reproduction. Two things to keep in mind before you reach for it: it is a raw base checkpoint, so it will not reliably follow instructions until someone post-trains it, and the FP8 weights alone are roughly 288GB — this is a data-center model, not a laptop download. For comparable fast, cheap models aimed at the same workloads, see [DeepSeek V4.1 Flash 1017](/ai-tools/deepseek-v4-1-flash-1017) and [MiniMax M3.1 Flash](/ai-tools/minimax-m3-1-flash).
The boundary that matters for content work: an ERNIE-5.5-Flash-class model, once post-trained, reads and writes text. It generates no image, video frame, caption overlay, carousel, or post, it keeps no persona's face consistent, and it publishes nowhere. Its output is words.
Here is the useful way to think about ERNIE 5.5 Flash Base for content: it represents the drafting layer getting cheaper, not a tool you can open today. It is a base checkpoint, so until it is post-trained and hosted it produces no usable prompt-following text — and even when a fast, cheap ERNIE-derived API arrives, what it gives you is still just words, in bulk. The failure mode of cheap bulk text is always the same: you end up with thirty caption angles, a month of hooks, five script drafts, and a stack of outlines, and a pile of words is nowhere near a feed. Turning that into finished, published content is exactly what [Kompozy](/) does, and it is built to take a whole batch as input rather than one draft at a time.
What makes this bridge durable is that Kompozy is model-agnostic about the drafting layer. Whichever cheap-inference model wins — ERNIE 5.5 Flash, DeepSeek, or the next one — the finishing engine does not change: hand Kompozy the batch and it fans each piece into formats the model cannot render. A script becomes a captioned [Persona Shorts](/glossary/persona-shorts) avatar video, a listicle outline becomes a brand-exact [Carousel](/glossary/hyperframes), the standout line becomes a [Quote Graphic](/glossary/persona-tweet), a long draft becomes a Blog Article and an Email Newsletter — one [Persona Brief](/glossary/persona-brief) governs voice and banned phrases across the whole batch, so output that came from a raw model still reads as one brand. Then [Autopilot](/glossary/autopilot) reframes each asset to 9:16, 1:1, and 16:9 and schedules it across the eight social platforms plus blog and email behind a per-post review gate. One precise caveat so this stays honest: Kompozy's own copy generation runs on Claude and OpenAI, so ERNIE 5.5 Flash Base is an upstream drafting foundation, not a plug-in you connect to Kompozy.
ERNIE 5.5 Flash Base is Baidu's open-weight base language model, released November 5, 2026 — a sparse mixture-of-experts model with 280B total parameters and about 14B activated per token, pre-trained on 33.5 trillion tokens. It is a pre-training checkpoint released before instruction tuning, so it is a foundation to fine-tune rather than a finished chat model.
Not directly. As a base checkpoint it has no instruction tuning and will not reliably follow a prompt until post-trained, and running the weights needs data-center GPUs. Even post-trained it produces only text. To turn drafts into finished posts and publish them, you bring that text into a content engine like Kompozy.
The weights are published openly (no license fee from this release) on Hugging Face and ModelScope under the ERNIE 5.5 Flash Base Model License Agreement v1.0. The real cost is running it — the FP8 weights are roughly 288GB, so it needs data-center hardware. Any cheap per-token price you eventually see would come from a provider hosting a post-trained version, not from Baidu directly.
No. It is a text base model. Even once post-trained it reads and writes text and generates no images, video, or audio. For visual content you pair a language model with tools that render images and video — which is what Kompozy does, using gpt-image, Gemini, and HeyGen.
Over-generate cheaply on a fast model — many caption, hook, script, and outline variants — then bring the winners into Kompozy to generate carousels, persona/avatar video, quote cards, infographics, blogs, and newsletters in your brand voice and publish across the eight social platforms plus blog and email.