ERNIE 5.5 Flash Base: Baidu’s 280B Open-Weights Base Model (Specs, Benchmarks, Hardware, Licence)

Baidu has just put the weights of a 280-billion-parameter model on Hugging Face, and it is not the usual chatbot drop. ERNIE 5.5 Flash Base, released on 5 November 2026, is a raw pre-training checkpoint: no instruction tuning, no RLHF, no chat template. It is meant for teams who want to build their own model on top of a strong foundation. In Baidu’s own tests it beats DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base on most benchmarks while using only 14B active parameters per token. There is a catch, though, and it is the licence. This guide covers the full specs, every benchmark Baidu published, the exact download and serve commands, the hardware you really need, what the licence allows (and where it does not apply at all), and what is still missing.

What Is ERNIE 5.5 Flash Base?

ERNIE is Baidu’s family of large language models, the same family that powers the ERNIE Bot assistant in China. The history matters here, because Baidu has changed its open-weights policy twice in eighteen months:

• ERNIE 4.5 (30 June 2025): ten variants released under Apache 2.0, with no territory limits and no revenue gate.
• ERNIE 5.0 (November 2025): no weights released at all.
• ERNIE 5.5 Flash Base (5 November 2026): weights are back, but under a custom licence that is far narrower than Apache 2.0.

The “Base” part is the important word. Baidu describes it as the checkpoint taken at the end of pre-training, before any mid-training or post-training, offered as “a common starting point for domain adaptation, instruction tuning and reinforcement-learning research”. In practice that means it is excellent at continuing text, few-shot prompting and being fine-tuned, but it will not follow instructions like ChatGPT or Claude. If you type “Write me a poem”, it may well continue with more questions instead of a poem.

Why it matters: open base models of this size are rare. Most labs only release their instruct models, which already carry the lab’s own alignment choices. A clean base lets researchers and companies run their own SFT and RL recipes from scratch, and Baidu’s numbers suggest this one is the strongest base in the 280-310B MoE class right now.

Full Specs and Architecture

Every figure below comes from the official model card on Hugging Face.

Three design choices stand out. First, ECHO reuses the final layer as a built-in draft model for speculative decoding, instead of bolting on a separate multi-token-prediction module. Second, the hybrid attention keeps most layers on a heavily compressed cache, which is how Baidu makes a 1M-token context affordable. Third, the right-to-left number tokenizer is an old trick for better arithmetic, and it may explain the strong MGSM and MathBench scores below. Baidu has published separate papers on its Virtual Head Attention, Multi-View MoE and ECHO for anyone who wants the maths.

Benchmarks: All 30 Scores Baidu Published

These are vendor-reported numbers. Baidu says every model was run through the same pipeline, prompts and few-shot settings, but no independent lab has reproduced them yet. Bold marks the best score in each row.

How to read this: the biggest gaps are on MMLU-Pro (+9.2 over DeepSeek’s base), HumanEval+ (+12.8) and LiveCodeBench (+8.2). Those are large margins for models with nearly the same active size. DeepSeek still wins on ChineseSimpleQA, HELMET-Rerank and InfiniteBench, and Tencent’s Hy3 preview wins FullStackBench and BBH with chain-of-thought. Remember these are base-model scores. They tell you how good a starting point it is, not how a finished chat model built on it will feel.

Hardware: What You Need to Run It

Baidu does not publish a minimum spec, so the table below is my estimate from the file sizes and the launch commands on the model card (which use 4-way tensor or data parallelism). Weights alone need the memory shown, plus room for the KV cache.

So, honestly: this is not a model for a home rig today. If you want to try it, rent a node by the hour. My cheapest cloud GPU guide lists current H100, H200 and B200 hourly prices on RunPod, Vast.ai and Lambda. For a general sense of how parameter count turns into memory, see how much VRAM you need for AI.

How to Download and Serve It (vLLM and SGLang)

Before you download 288 GB, check two things: you are outside the EU, UK and South Korea (the licence does not cover those regions), and you have accepted the licence on the Hugging Face page.

Step 1: Download the weights

pip install -U "huggingface_hub[cli]"
huggingface-cli login
# FP8 (about 288 GB)
huggingface-cli download baidu/ERNIE-5.5-Flash-Base --local-dir ./ERNIE-5.5-Flash-Base
# or BF16 (about 573 GB)
huggingface-cli download baidu/ERNIE-5.5-Flash-Base-BF16 --local-dir ./ERNIE-5.5-Flash-Base-BF16

ModelScope hosts mirrors of both repos if Hugging Face is slow in your region.

Step 2a: Serve with vLLM (Baidu’s recipe)

Plain vLLM needs Baidu’s plugin for this architecture. This is the deployment recipe from the model card:

uv venv --python 3.12 ernie5-venv
source ernie5-venv/bin/activate
uv pip install vllm==0.32.0
uv pip install git+https://github.com/baidu/vLLM-ERNIE.git
vllm serve baidu/ERNIE-5.5-Flash-Base \
  --kv-cache-dtype fp8 \
  --enable-expert-parallel \
  --data-parallel-size 4 \
  --max-model-len 131072 \
  --port 30000 \
  --moe-backend humming

Step 2b: Serve with SGLang (Docker)

docker run -d -it --name ernie55 --gpus all --network host --ipc host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  lmsysorg/sglang:dev-ernie-55-flash

# inside the container
sglang serve \
  --model-path baidu/ERNIE-5.5-Flash-Base \
  --tp 4 --attn-dp-size 4 --context-length 131072 \
  --max-running-requests 1024 --host 0.0.0.0 --port 30000

Step 3: Send a completion request

Use the /v1/completions endpoint, not /v1/chat/completions. A base model has no chat template.

curl http://localhost:30000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "baidu/ERNIE-5.5-Flash-Base", "prompt": "Large language models are", "max_tokens": 256, "temperature": 0}'

Step 4: Prompt it like a base model

To get useful work out of a base model, show it the pattern you want. A few examples work far better than an instruction:

Q: Convert 3 km to miles.
A: 1.86 miles
Q: Convert 10 kg to pounds.
A: 22.05 pounds
Q: Convert 25 C to Fahrenheit.
A:

Transformers also works for quick tests (AutoModelForCausalLM.from_pretrained("baidu/ERNIE-5.5-Flash-Base", device_map="auto") after pip install -U transformers accelerate), but it is much slower than vLLM or SGLang for a model this size.

Fine-Tuning: The Real Point of a Base Model

Baidu built this release for people who want to make their own model. The options today:

• Full SFT with NVIDIA Megatron Bridge: NVIDIA ships an ERNIE 5.5 fine-tuning example covering data prep, training and export back to Hugging Face format. Its reference setup is 16 nodes of 4x GB300. It is validated on Blackwell (GB200, GB300); Hopper is supported but not validated.
• Parameter-efficient tuning (LoRA and friends): the realistic route for most teams, but there is no official LoRA recipe yet. Expect community recipes once Hugging Face PEFT and the training frameworks add the new architecture.
• Long-context tuning: the config allows 1M tokens, but pre-training stopped at 128K, so long-context behaviour beyond 128K needs your own tuning and testing.

One licence rule matters a lot here: you may not use the model or its outputs to improve any other AI model. That rules out using ERNIE to generate synthetic training data or distil into, say, a Qwen or Llama model.

The Licence Explained in Plain English

This is a custom licence, not open source in the OSI sense. Here is what the ERNIE 5.5 Flash Base Model License Agreement v1.0 says, section by section. (This is a summary, not legal advice; read the full LICENSE file before you deploy.)

ERNIE 5.5 Flash Base vs Other Open Models

Who should use which: if you just want a strong open model to chat with or build an app on today, ERNIE 5.5 Flash Base is the wrong choice; an instruct model such as GLM-5.3 Flash or DeepSeek V4.1 Flash (see my DeepSeek API pricing guide) is ready to go. If you are a lab or company outside Europe that wants a clean, high-scoring base for your own post-training, ERNIE 5.5 Flash Base now looks like the strongest candidate in its class on paper. If you are in the EU or UK, you cannot use it at all. For newer models that you can actually run at home, my best Ollama models list and best GPU for local LLMs guide are the better starting points. Tencent’s own next model is also coming; see my Tencent Hy4 coverage.

Troubleshooting

“Model architecture ernie5_5_moe is not supported”

Your engine is too old or lacks the plugin. For vLLM, install exactly vllm==0.32.0 plus git+https://github.com/baidu/vLLM-ERNIE.git in a fresh environment. For SGLang, use the lmsysorg/sglang:dev-ernie-55-flash image rather than a stock release. For transformers, run pip install -U transformers accelerate. llama.cpp, Ollama, LM Studio and mlx-lm cannot load this architecture yet; there is nothing to fix on your side until they add support.

The output rambles or ignores my request

That is expected from a base model. Use /v1/completions, write few-shot prompts, set temperature to 0 for factual tasks, and add a stop sequence (for example "\nQ:") so it stops after one answer. Chat-style system prompts will not work until Baidu ships the post-trained version.

/v1/chat/completions returns an error about a chat template

The base checkpoint has no chat template. Switch to the completions endpoint.

Out of memory on 4x 80 GB GPUs

The FP8 weights take about 288 GB of your 320 GB. Keep --kv-cache-dtype fp8, drop --max-model-len from 131072 to 32768 or lower, lower --max-running-requests in SGLang, or move to 8x H100 or 4x H200.

Hugging Face download refused or gated

Log in with huggingface-cli login, accept the licence on the model page, and make sure you are not in a region excluded by the licence. Try the ModelScope mirror if the transfer keeps failing.

What We Still Don’t Know

• The instruct model: Baidu says a post-trained version is coming, with no date or licence details.
• Independent benchmarks: no third party has reproduced Baidu’s numbers yet.
• 1M-token quality: enabled in config, but trained only to 128K.
• Hosted APIs: no router or cloud lists it at launch and no price is published.
• Quantised builds: no GGUF, AWQ or MLX versions exist yet.

My Take

On the numbers alone, this is a big release. A base model that leads its size class on MMLU-Pro and coding by 8 to 13 points, at only 14B active parameters, would normally be the obvious new foundation for every open fine-tune. The architecture is interesting too: ECHO’s built-in drafting and the compressed attention look built for cheap long-context serving. But the licence pulls in the other direction. Shutting out the EU, UK and South Korea, adding a revenue gate, and banning outputs from improving other models makes it a poor fit for the open community’s usual mix-and-match workflows, and a step back from ERNIE 4.5’s Apache 2.0. I’d treat ERNIE 5.5 Flash Base as an excellent research and in-house base for teams in the licensed territory, and I’ll be watching for the instruct version and the first independent evaluations before calling it the new default.

Frequently Asked Questions

Is ERNIE 5.5 Flash Base open source?

The weights are downloadable, but the licence is a custom Baidu agreement, not an OSI open-source licence. It limits regions, gates large commercial users and restricts using outputs to train other models.

Can I use ERNIE 5.5 Flash Base in the EU or UK?

No. The licence Territory is worldwide except the European Union, the United Kingdom and South Korea, and use or display outside the Territory is not allowed.

Can I run ERNIE 5.5 Flash Base locally with Ollama or LM Studio?

Not yet. llama.cpp and MLX do not support the architecture, and no GGUF exists. Even a 4-bit build would be around 160 GB, beyond consumer GPUs.

How much VRAM does ERNIE 5.5 Flash Base need?

The FP8 weights are about 288 GB and BF16 about 573 GB, plus KV cache. A 4x H200 or 8x H100 node is a sensible minimum for FP8 serving.

Is ERNIE 5.5 Flash Base a chatbot?

No. It is a pre-training checkpoint that continues text. Baidu says a post-trained version that follows instructions will come later.

Is ERNIE 5.5 Flash Base better than DeepSeek V4 Flash?

Compared base to base, Baidu’s table shows ERNIE ahead on 26 of 30 benchmarks, including MMLU-Pro (78.4 vs 69.2) and HumanEval+ (82.3 vs 69.5). These are vendor-reported and compare base checkpoints, not the chat models.

Is it free for commercial use?

Yes up to US$50 million in annual related revenue, if you notify Baidu once before launch and display “Powered by ERNIE”. Above that you need Baidu’s written authorisation.

Sources

• Hugging Face: baidu/ERNIE-5.5-Flash-Base model card
• ERNIE 5.5 Flash Base Model License Agreement v1.0
• Baidu ERNIE blog: ERNIE 5.5 Flash Base
• The Decoder: Baidu releases ERNIE 5.5 Flash Base
• ModelFit: file sizes and local engine support
• OrcaRouter: licence analysis and ERNIE 5.5 Flash vs GLM-5.3 Flash
• DEV Community: fine-tuning notes (Megatron Bridge recipe)