Baidu Opens the Weights on ERNIE 5.5 Flash Base, With Strings Attached

Baidu ships ERNIE 5.5 Flash Base as a raw starting point

The ERNIE 5.5 Flash Base model is now available to download. Baidu announced the release on November 5, 2026, publishing the weights for a text-only mixture-of-experts model with 280 billion total parameters and 14 billion activated per token. This is a base model, not a finished chatbot. The checkpoint stops at the end of pre-training, before any mid-training or post-training, which means it is a starting point for other people to build on rather than something you chat with out of the box.

That distinction is the whole story. A base model gives researchers a fixed, shared checkpoint for domain adaptation, instruction tuning and reinforcement-learning experiments, without anyone having to repeat Baidu's pre-training run. Baidu says a post-trained ERNIE 5.5 Flash model will follow, but for now the base version is what you get.

What is inside the release

The model card lists 43 layers, a hidden size of 4,096 and 64 attention heads. Downloads come in FP8 and BF16 safetensors formats, and Baidu documents production serving through SGLang and vLLM using the completions API, which matches the checkpoint's base-model status.

Baidu also published FlashMask, a repository of attention kernels for Hopper and Blackwell GPUs, with Blackwell configurations supporting head dimensions up to 576 for query and key and 512 for value. That is the kind of plumbing detail that matters to the people this release is actually for: engineers who want to run the model fast on modern Nvidia hardware.

On benchmarks, Baidu evaluated ERNIE against the publicly released DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base checkpoints using the same harness, prompts and few-shot examples, plus the same LLM judge for free-form question answering. Same-harness comparisons are the honest way to do this, and they make the results more useful than a cherry-picked leaderboard. The numbers still come from Baidu, so treat them as a first signal rather than a verdict.

The license is the catch

The ERNIE 5.5 Flash Base Model License Agreement v1.0 imposes territorial and commercial conditions. Its territory excludes the European Union, the United Kingdom and South Korea, and use, modification and distribution outside that territory are not authorized. That is a sharp limit for a release that otherwise looks open. A research team in Berlin or Seoul cannot legally use these weights under the license as written, which cuts out a large share of the global research community.

This is part of a pattern. Chinese labs have been shipping capable open-weight models at a fast pace this month, including Tencent Hy4 and Meituan LongCat-2.5, and the licenses attached to them increasingly shape who can actually build on the work. Weights you can download are only half the story. The license decides the rest.

Who should care about this release

If you fine-tune models for a specific domain, a fresh 280B-parameter base checkpoint with a fixed baseline for controlled comparisons is genuinely useful. It lets you test training methods against a known starting point instead of guessing whether your improvements came from the method or the model. That is the audience Baidu is speaking to, and the release is built for them.

If you just want a capable model to chat with, wait for the post-trained ERNIE 5.5 Flash. A base model will ramble, repeat itself and happily complete your prompt in the wrong direction, because that is what base models do. They predict text, not answer questions.

The bigger signal is the pace. November has already delivered a run of large open-weight releases, and ERNIE 5.5 Flash Base adds another heavyweight to the pile. The open question, as with Mistral Large 4, is whether independent evaluations confirm Baidu's numbers and whether the license restrictions push researchers toward friendlier alternatives. Download it, read the license first, and benchmark it yourself before betting a project on it.