What did Baidu release in ERNIE 5.5 Flash Base?
Baidu released ERNIE 5.5 Flash Base on November 5, 2026. It is a text-only sparse mixture-of-experts model with 280 billion total parameters. Fourteen billion of those parameters are active for each token.
Baidu published weights in FP8 and BF16 formats on Hugging Face and ModelScope. The checkpoint is the model at the end of pre-training. It has had no mid-training or post-training, so it is not an instruction-tuned assistant.
The company trained it from scratch on 33.5 trillion tokens. Training used the Muon optimizer and FP8 mixed precision. The training context reached 128,000 tokens, and the released configuration supports up to one million tokens for fine-tuning and inference.
The architecture has 43 Transformer layers, a hidden size of 4,096 and 64 attention heads. The first layer uses a dense feed-forward block. The other 42 use a latent mixture-of-experts block with 512 routed experts. Each token is routed to the top 10 experts plus one shared expert.
Attention is hybrid by design. Thirty-seven layers use Heavily Compressed Attention, which caches one entry for every 128 tokens and attends to a short window of recent tokens. Six layers keep token-level access through Multi-head Latent Attention with sparse attention. Baidu says the split keeps long-context costs down while preserving accuracy.
Smaller details point at the same efficiency goal. The tokenizer has a vocabulary of about 200,000 entries and groups digits from right to left in threes. Baidu replaced its multi-token prediction module with ECHO, which reuses the final layer to predict future tokens. A custom attention kernel, FlashMask-4, is up to 175% faster in the forward pass than FlashAttention-4 MaskMod on Nvidia SM100 hardware, according to Baidu.
Routing also differs from a standard mixture-of-experts layer. Baidu uses several independently initialized routers that score experts together instead of relying on a single router. The company says this reduces bad selections, with the largest gains in the middle and deeper layers. The number of activated experts and the communication volume stay the same.
How does ERNIE 5.5 Flash Base compare with DeepSeek and other open models?
Baidu compared the model with three open base checkpoints of similar size. They were DeepSeek-V4-Flash-Base at 284 billion total and 13 billion active parameters, Hy3-preview-Base at 295 billion and 21 billion, and MiMo-V2.5-Base at 310 billion and 15 billion.
All four ran through the same evaluation harness, prompts, few-shot examples and LLM judge. Baidu published 30 benchmark rows. ERNIE 5.5 Flash Base has the best score in 24 of them.
The suite covers familiar ground. MMLU and MMLU-Pro test broad exam knowledge. GPQA tests graduate-level science questions. HumanEval+, MBPP+ and LiveCodeBench measure code generation. HELMET, RepoQA and LongBench-v2 probe retrieval and reasoning over long inputs.
The exam and knowledge results show a consistent lead. It scores 93.3 on MMLU against 89.4 for DeepSeek. It scores 78.4 on MMLU-Pro against 69.2. On GPQA it reaches 60.2 against 55.9, and on SimpleQA it reaches 34.1 against 29.1.
Coding shows a wider gap. ERNIE scores 82.3 on HumanEval+, 81.5 on MBPP+ and 40.1 on LiveCodeBench. DeepSeek scores 69.5, 67.2 and 31.9 on the same tests. On the long-context suite, ERNIE reaches 71.5 on HELMET-RAG, 91.7 on RepoQA and 48.7 on LongBench-v2.
It does not win every row. DeepSeek leads on ChineseSimpleQA at 75.3 against 73.0 and on INCLUDE-lite at 79.8 against 79.5. DeepSeek also leads on HELMET-Rerank at 60.0 against 56.2 and InfiniteBench at 48.0 against 46.5. Hy3-preview-Base leads on chain-of-thought BBH at 90.0 against 89.2 and FullStackBench at 56.7 against 54.6.
The active-parameter comparison is the practical takeaway. ERNIE activates 14 billion parameters per token, more than DeepSeek's 13 billion but less than Hy3's 21 billion and MiMo's 15 billion. It still leads most rows in Baidu's table, which is the efficiency claim the release rests on.
These are vendor-run numbers. Baidu ran all four checkpoints itself, so the table compares architectures under one setup but is not a neutral audit. Base-model scores measure next-token prediction after pre-training. They do not show how helpful the model becomes after alignment, and they are sensitive to prompt and judge choices. No independent replication of the full table was available when this article was published.
What does the ERNIE 5.5 Flash Base license allow?
The weights ship under the ERNIE 5.5 Flash Base Model License Agreement v1.0. It is a custom license dated November 5, 2026, and it is not an OSI-approved open-source license. Baidu describes the release as open, and open weights is the more accurate description.
The license grants the right to run, reproduce and modify the materials, to continue training and fine-tune, to create derivative models and to distribute them. The territory clause carries the most weight. The license defines the territory as all areas worldwide other than the European Union, the United Kingdom and South Korea. Rights are granted solely inside that territory.
Section 5.3 goes further. It bars use, reproduction, modification, distribution or display of the weights, derivatives and outputs outside the territory. That covers research use as well as commercial use. Developers in the three excluded regions receive no license grant at all.
Commercial terms are tiered by revenue. Non-commercial use is free. Commercial use is also free while annual revenue from related products or services stays at or below US$50 million. The operator must send a one-time notice to Baidu and display Powered by ERNIE on its interface, website or documentation.
Above US$50 million, the rules change. The operator must obtain separate written authorization before starting or continuing commercial use, and fees are set in a separate agreement. Revenue is assessed continuously, and Baidu can request a compliance statement once a year.
Two other clauses matter for model builders. Section 5.5 forbids using the materials, derivatives or outputs to improve any AI model other than ERNIE 5.5 Flash or its derivatives, which blocks distillation into a rival system. Section 6.6 terminates the license if the user sues Baidu over intellectual property.
The agreement also imports an acceptable-use policy and requires distributors to pass the same terms to recipients. Anyone offering a service that generates outputs must maintain safeguards against misuse and provide a way to report violations. Those duties travel downstream, so a company cannot loosen the terms for its own customers.
Baidu's release post does not explain the regional carve-out. The practical effect is clear. Teams elsewhere can fine-tune, modify and redistribute under the same terms. Teams inside the EU, UK or South Korea cannot rely on the license. Large companies outside those regions face a negotiation rather than a published price.
Why release a base checkpoint instead of a finished chatbot?
Baidu argues that a base checkpoint is a reusable asset. In its release post, it says different domains, reasoning tasks and agentic applications need different data, objectives and behaviours. Releasing weights after pre-training lets developers choose where to take the model.
The company also presents the checkpoint as a clean research baseline. It is a single snapshot taken before mid-training or post-training. Labs can compare fine-tuning and reinforcement-learning methods from the same starting point and measure what each one adds.
There is a cost argument too. Pre-training on 33.5 trillion tokens is expensive. Baidu says it hopes the release spares developers the cost of large-scale pre-training, so they can spend their resources on the capabilities they care about.
The trade-off is readiness. The model is a text-completion system, not a chat assistant. The model card tells users to keep the beginning-of-sequence token when tokenizing. Chat behaviour, refusal behaviour and tool use would all have to come from later training.
Baidu says a post-trained ERNIE 5.5 Flash will follow. The Decoder reported the same plan. Until that model ships, the release mainly serves teams with their own training pipelines. Businesses looking for a drop-in assistant have little to run.
Rivals publish base checkpoints too. DeepSeek, the Hy team and the MiMo team all supplied base versions for Baidu's comparison. That shows the format is established among the labs Baidu benchmarks against, even if it remains a niche product for most buyers.
What hardware does a 280B model need?
Memory is the first constraint. At one byte per parameter, FP8 weights need roughly 280 GB before any KV cache or runtime overhead. The BF16 version needs roughly twice that. Neither fits on a single 80 GB accelerator.
Baidu's model card gives deployment recipes for vLLM and SGLang. The vLLM example uses four-way data parallelism and a maximum model length of 131,072 tokens. The SGLang example uses four-way tensor parallelism, the same 131,072-token context and up to 1,024 concurrent requests. Both assume a multi-GPU server.
The vLLM path also needs a Baidu fork of the serving stack, installed over vLLM 0.32.0. That is an extra integration step compared with models that run on a stock release.
Baidu hosts the BF16 checkpoint in a separate Hugging Face repository. The model card lists no Hugging Face inference provider for either version, so self-hosting is the default path. That matters for teams that expected a hosted endpoint on release day.
Long context adds its own cost. Baidu says inference can reach one million tokens, but its published serving examples stop at 128K. The architecture is meant to make long prompts cheaper. Most layers cache one entry per 128 tokens, so the KV cache grows far more slowly than a dense model's would.
Six layers keep full token-level caches, and they use sparse attention to control the cost. Baidu says those layers are enough to keep long-context benchmark scores competitive.
Baidu does not publish an inference price for the base checkpoint. The license says the company is not committing to hosted inference, compute, training or support. Teams that want a managed endpoint need a third-party provider or the post-trained model's commercial terms.
How does this fit Baidu's open-weight strategy?
Baidu first released open weights in June 2025, when it published 10 variants of ERNIE 4.5 on Hugging Face from 0.3 billion to 424 billion parameters. The South China Morning Post called the move a U-turn. It noted that founder Robin Li had previously argued closed models would stay more powerful.
The family has kept moving since. ERNIE 5.0, a 2.4 trillion-parameter multimodal model, was announced in February 2026. ERNIE 5.1 followed in May 2026. Baidu also says an ERNIE 5.1 preview topped LMArena's text leaderboard among Chinese models in April 2026. ERNIE 5.5 Flash Base is the first checkpoint in the 5.5 series.
The comparison set is deliberately narrow. Baidu tested three base models, all between 284 billion and 310 billion total parameters. Larger open models and instruction-tuned systems sit outside the table. Baidu chose the models and the tests.
Open weights serve several purposes for a lab in Baidu's position. They attract fine-tuning work, tooling support and researchers who might otherwise build on a rival. They also build goodwill with developers who cannot afford frontier API prices.
The license complicates that strategy. A territorial ban and a revenue threshold create friction for exactly the well-funded companies that could turn a base model into a commercial product. Baidu appears to be trading broad adoption for control over large deployments.
The carve-out also narrows the contributor pool. European and British labs cannot publish derivatives under this license, even though they can read the release and the research. That limits the fine-tuning work Baidu can point to later.
What should developers and buyers watch next?
The post-trained model is the next milestone. Baidu says ERNIE 5.5 Flash will follow the base release. That version will matter more for chatbots, coding assistants and agents. Its license and API price will decide whether the family gets commercial traction.
Independent evaluation is the second thing to watch. Baidu's 24-of-30 result comes from its own harness. Third-party leaderboards and replication studies will show whether the gap over DeepSeek and MiMo holds outside the company's setup.
Adoption signals are worth tracking. The Hugging Face repository had 526 downloads in the previous 30 days and 202 likes when checked on November 6. That is modest for a model of this size. Derivative fine-tunes and vendor integrations would be a stronger sign.
Licensing details may also move. Baidu could clarify why the EU, UK and South Korea are excluded, or offer separate terms for those markets. Enterprises above the US$50 million threshold will negotiate privately, so public disclosures may be sparse.
For developers, the decision is straightforward. Teams outside the excluded regions can test the weights, fine-tune them and ship with attribution. Teams inside those regions can read the research but have no license to use the model.