Baidu ERNIE 5.5 Flash Base: Free 280B Weights, $50M License Catch

Baidu has put one of the largest freely downloadable AI models of the year on Hugging Face. ERNIE 5.5 Flash Base, released on November 5, 2026, is a 280-billion-parameter language model that anyone can download, with 14 billion parameters active per token. Baidu's own benchmark table puts it ahead of DeepSeek's V4 Flash base model on 24 of 30 tests. That is a serious result for a checkpoint that costs nothing to download.

The headline number is not the whole story. This is a base model, not a chatbot. It has no instruction tuning, no chat template and no safety alignment, so it cannot answer questions out of the box the way ChatGPT or Claude can. And the license that comes with it is stricter than the one Baidu used last time: use is barred in the European Union, the United Kingdom and South Korea, and commercial use above $50 million in related revenue needs Baidu's written permission. The model is free. The terms are not.

What ERNIE 5.5 Flash Base actually is

ERNIE 5.5 Flash Base is the raw checkpoint at the end of pre-training, before any mid-training or post-training. Baidu trained it from scratch on 33.5 trillion tokens and then stopped, releasing the weights so other teams can decide what happens next: domain adaptation, instruction tuning, reinforcement learning, or controlled comparisons of training methods from a shared starting point. A post-trained ERNIE 5.5 Flash will follow, according to Baidu, but there is no date for it yet.

That makes it a research artifact first and a product second. It is also Baidu's first open-weight release since the ERNIE 4.5 series in June 2025, according to reporting on the launch; the ERNIE 5.0 and 5.1 models were never open-sourced. The gap makes this release the first chance in roughly 16 months for outside teams to work directly with Baidu's latest architecture.

The model is text-only and bilingual in practice, with English and Chinese listed on its model card. It is available in two weight formats on Hugging Face, FP8 and BF16, and on ModelScope, Baidu's domestic hub. The BF16 variant is a separate download under the same license.

The architecture, in plain terms

The headline design choice is efficiency. ERNIE 5.5 Flash Base has 280 billion total parameters but activates only 14 billion of them for each token, using a sparse mixture-of-experts layout: 512 routed experts per layer, of which 10 are selected for each token, plus one shared expert. The model has 43 layers, a hidden size of 4,096, 64 attention heads and a vocabulary of 201,216 tokens.

The more unusual part is the attention system. Thirty-seven of the 43 layers use what Baidu calls Heavily Compressed Attention: each layer's cache holds one entry for every 128-token block, plus a short window of recent uncompressed tokens. The remaining six layers use Multi-head Latent Attention with dynamic sparse attention, which keeps token-level access to the full context and picks out the highest-scoring earlier positions. Baidu says this keeps per-query core-attention cost constant as context grows. The context used during training was extended gradually to 128K tokens; fine-tuning and inference support up to 1 million tokens.

Two other pieces are worth knowing about. The model uses several independently initialized routing networks instead of one, because Baidu says a single router's scores correlate weakly with expert usefulness in deeper layers. And a method called ECHO reuses the backbone's final layer to predict several future tokens at once, which lets the same layer act as a draft model during speculative decoding instead of needing a separate draft checkpoint. Training used the Muon optimizer and FP8 mixed precision, and Baidu documents attention kernels, under the name FlashMask, that it says run up to 175% faster forward and 74% faster backward than FlashAttention-4 MaskMod on Blackwell hardware.

Baidu's benchmark table: strong, but not a clean sweep

Baidu evaluated ERNIE 5.5 Flash Base against three publicly released base checkpoints, DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base, using the same harness, prompts, few-shot examples and LLM judge for free-form answers. Counting Baidu's own table, ERNIE comes first on 24 of the 30 benchmarks. The wins are widest where the pre-training data lines up with the test: the corpus includes repository-level software records such as commits, issues, pull requests, reviews and CI-verified fixes, and the code scores show it.

Every number in that table is Baidu's own, run on Baidu's harness, and none of it has been independently reproduced. That matters more than usual here because the comparison set is Baidu's choice: three rival base checkpoints, all released recently, all measured the same way. The table is still useful, because it shows where the model is strong and where it is not. The six tests where ERNIE does not lead are ChineseSimpleQA, where DeepSeek scores 75.3 against ERNIE's 73.0; BBH with chain-of-thought, where Hy3 scores 90.0 against 89.2; INCLUDE-lite, where DeepSeek leads 79.8 to 79.5; FullStackBench, where Hy3 leads 56.7 to 54.6; HELMET-Rerank, where DeepSeek leads 60.0 to 56.2; and InfiniteBench, where DeepSeek leads 48.0 to 46.5.

The pattern is consistent: ERNIE leads on exams, knowledge and code, and trails on factual retrieval, ranking and sustained performance over very long inputs. If your use case is long-document question answering or retrieval, the two InfiniteBench and HELMET results are a warning that the headline 24-of-30 figure does not cover. For context on the rivals, DeepSeek's V4.1 Flash is a different, newer release from the same lab, priced at $0.15 per million input tokens on DeepSeek's own API, which we covered separately.

The license is the real story

ERNIE 5.5 Flash Base ships under the ERNIE 5.5 Flash Base Model License Agreement v1.0, a custom license Baidu wrote for this release. That is a step back from ERNIE 4.5, which used the Apache 2.0 license. Apache 2.0 is permissive and familiar; this agreement is neither, and it changes who can actually use the model.

First, the territory. The license excludes the European Union, the United Kingdom and South Korea, and activities including use, modification and distribution outside the licensed territory are not authorized. A team based in London, Berlin or Seoul cannot use this model under its terms, whatever the download page says. That is an unusual restriction for a major open-weight release, and it will shape adoption more than any benchmark number: the US and Chinese research communities can work with it, while three of the world's largest AI markets cannot.

Second, the commercial line. Non-commercial use is free of model license charges. Commercial use at or below $50 million in annual revenue from related products or services requires a one-time notice to Baidu and prominent "Powered by ERNIE" attribution. Above that threshold, users need Baidu's separate prior written authorization. The $50 million figure is measured on revenue from related products or services over the prior 12 consecutive months, so a company that grows into the threshold has to go back to Baidu before it keeps shipping.

Third, the distillation ban. The agreement bars use of the licensed materials, derivative models or their outputs to improve an unrelated AI or machine-learning system. That restriction covers distillation and synthetic-data workflows, which the license addresses directly when it defines derivative models and limits on output use. Teams that train smaller models on the outputs of larger ones, a common way to build cheap specialized models, need to read this clause before they start. It does not ban fine-tuning ERNIE itself; it bans using ERNIE to make a different model better.

None of this makes the release dishonest. Baidu is upfront about the terms on the model card, and the license is there to read before downloading. But it does mean "open weights" and "open source" are not the same thing here. The weights are downloadable; the permission to use them is conditional, territorial and revocable in the commercial case above $50 million. Anyone comparing this release with fully permissive ones should keep that distinction in mind. Tencent's Hy4, released a day earlier, also carries a weights-pending caveat of its own, and Kimi K3.1 Pro's weights landed on Hugging Face the same week under a separate license, so the open-weight landscape right now is full of releases that look free until the fine print.

What it costs to actually run it

The download itself is free, and there is no API price because Baidu is not selling hosted access to this checkpoint. The real cost is hardware. Only 14 billion parameters activate per token, but a conventional deployment still has to hold the full 280-billion-parameter pool in memory. The FP8 weights run to roughly 280GB, and the BF16 version roughly double that, before the KV cache, activations and runtime overhead. That is arithmetic from the parameter count, not a published figure, but it sets the scale: this is a multi-GPU model, not a laptop download.

Baidu documents production serving with SGLang and vLLM through the completions API, which matches the checkpoint's base-model status. The vLLM example pins version 0.32.0 and installs a dedicated plugin from Baidu's vLLM-ERNIE repository. NVIDIA's Megatron Bridge adds a supervised fine-tuning workflow covering data preparation, training and export back to Hugging Face format; its full-model recipe uses 16 nodes with four GB300 GPUs per node, and NVIDIA validates the recipes on Blackwell GB200 and GB300 systems. The example serving configuration caps requests at 131,072 tokens, so exposing the advertised 1,048,576-token limit needs a matching setting and enough memory for the larger cache.

There is no published price for any of this because there is nothing to buy from Baidu. The cost is whatever your GPU cluster costs, plus the post-training work the checkpoint still needs. That is the honest way to read "free" here: free weights, expensive everything else.

Who should download it, and who should skip it

Research groups get the most from this release. A fixed checkpoint at the end of pre-training lets teams compare instruction tuning, reinforcement learning, data mixtures and alignment methods from the same starting point, without first undoing someone else's chat tuning. Architecture researchers can study the compressed attention, multi-router expert selection and ECHO speculative-decoding head independently. Domain teams that need a model shaped to their own data, output formats and refusal policies, and that have the compute for post-training, get a strong starting point with unusually good code scores.

Everyone else should pause. If you are based in the EU, UK or South Korea, the license says no. If you want a chatbot, this is the wrong download: it has no chat template, no assistant behavior and no refusal policy, and turning it into one is a project, not a prompt. If your product's related revenue is anywhere near $50 million, get legal advice before building on it, because the authorization requirement sits above that line. And if you plan to distill it into a smaller model, the license expressly restricts that use.

Smaller teams with a single GPU should look at the BF16 and FP8 sizes above and do the math before downloading hundreds of gigabytes. The FP8 checkpoint is the practical choice for serving, and even that needs multi-GPU infrastructure or sharding and offloading, which trades memory for latency.

Before you download: five checks

Read the license first, not the benchmark table. The territory exclusion and the $50 million line are the two clauses most likely to surprise a team after it has already invested in fine-tuning. Check that your use case fits inside them, and that your revenue, if any, is measured the way Baidu measures it.

Confirm the format you need. FP8 is the smaller download and the one Baidu's serving examples use; BF16 is roughly twice the size. Both are safetensors, and the model expects a BOS token at the start of every input, so keep add_special_tokens set to true when tokenizing.

Do not judge it on the chat benchmarks. Base-model scores measure pre-training, not assistant behavior. The MMLU and HumanEval+ numbers say the model learned a lot; they say nothing about how it will behave after your instruction tuning, which is the part you still have to build.

Budget for the long context separately. The 1-million-token figure describes fine-tuning and inference support, not a guarantee of good recall across the whole window. The InfiniteBench result, where DeepSeek leads, is the relevant evidence, and retrieval quality across very long inputs stays workload-dependent.

Watch for the post-trained version. Baidu says a post-trained ERNIE 5.5 Flash will follow this base release. If you do not need a research starting point, waiting for that version is the cheaper path: it should arrive with the chat behavior and alignment this checkpoint deliberately lacks.

Frequently asked questions

Is ERNIE 5.5 Flash Base free to use?

The weights are free to download, and non-commercial use carries no model license charge. Commercial use up to $50 million in annual revenue from related products or services needs a one-time notice and "Powered by ERNIE" attribution; above that, Baidu's separate written authorization is required. Use in the EU, UK and South Korea is not authorized under the license.

Can I run ERNIE 5.5 Flash Base on my own computer?

Not on a normal desktop. The FP8 weights run to roughly 280GB and the BF16 version roughly double that, before the KV cache and runtime overhead, so serving needs multi-GPU infrastructure or sharding and offloading. Baidu documents serving with SGLang and vLLM. A laptop is not a realistic target for the full model.

Is ERNIE 5.5 Flash Base better than DeepSeek V4 Flash?

On Baidu's own benchmark table, ERNIE leads on 24 of 30 tests, including MMLU at 93.3% against 89.4% and HumanEval+ at 82.3% against 69.5%. DeepSeek leads on six tests, including InfiniteBench and ChineseSimpleQA. All scores are Baidu's, run on Baidu's harness, and have not been independently reproduced, so treat the table as Baidu's claim, not a settled ranking.

Can I use ERNIE 5.5 Flash Base to train another model?

Not to improve an unrelated model. The license bars using the licensed materials, derivative models or their outputs to improve an unrelated AI or machine-learning system, which covers distillation and synthetic-data workflows aimed at a different model. Fine-tuning ERNIE itself is the intended use.

When will the post-trained ERNIE 5.5 Flash arrive?

Baidu says a post-trained ERNIE 5.5 Flash will follow the base version, but has not given a date. The base release is aimed at researchers and developers who want to do their own post-training; everyone else may find it cheaper to wait.

Baidu's release is a strong technical artifact with a license that narrows its audience on purpose. The benchmarks are real, the architecture is interesting, and the download is free. But the teams that can use it are the ones outside the EU, UK and South Korea, below $50 million in related revenue or willing to ask Baidu first, and equipped to do their own post-training. For everyone else, this is a model to watch, not a model to download. If that describes you, the model card and release notes below are the place to start.

Sources

• Baidu ERNIE Team, ERNIE 5.5 Flash Base release notes
• Baidu, ERNIE-5.5-Flash-Base model card on Hugging Face
• Baidu, ERNIE-5.5-Flash-Base-BF16 model card on Hugging Face
• Unite.AI, Baidu Releases ERNIE 5.5 Flash Base for Further Training
• KuCoin, Baidu open-sources the 280-billion-parameter ERNIE 5.5 Flash model after a 16-month gap
• AlphaSignal, Baidu Drops ERNIE 5.5 Flash, a Raw 280B Open Checkpoint Beating DeepSeek