ERNIE 5.5 Base weights arrive with South Korea outside the license territory

Baidu released ERNIE 5.5 Flash Base weights on November 5, providing a checkpoint for domain adaptation, instruction tuning and reinforcement learning. The release is taken at the end of pre-training, before mid-training or post-training.

The accompanying model license v1.0 excludes South Korea, the European Union and the United Kingdom from its permitted territory. Its grants of use, modification and distribution apply only inside that territory. The territorial restriction also covers derivative models and outputs.

The text model comes in FP8 and BF16 weight formats. Its model card lists 280 billion total parameters and 14 billion activated per token. Pre-training reached 128,000 tokens, while fine-tuning and inference support up to one million.

Baidu says most layers compress each 128-token span of history into one cached entry to reduce long-input storage, while a few retain token-level access. Training data includes repository changes, verified fixes, long-document retrieval and citation-grounded questions.

SGLang estimates weight memory at approximately 280GB for FP8 and 560GB for BF16. Its smallest recommended configurations with 288GB GPUs use two and four GPUs respectively. It marks its single-node FP8 configurations as verified; among BF16 recipes, only B300 is verified, with the others derived from memory budgets. For one-million-token requests on two B300 GPUs, it recommends single-digit concurrency. Long inputs need temporary computation space and a cache for input keys and values in addition to weights. It offers separate low-latency and high-throughput configurations.

Actual context capacity also depends on temporary memory. Baidu’s vLLM plugin documentation says processing one million tokens needs tens of additional gigabytes. Its 80GB-GPU configuration accommodates up to 262,144 tokens, but not one million. The plugin currently runs on NVIDIA GPUs, separating model-level context support from a particular serving setup.

NVIDIA’s Megatron Bridge example describes full-model supervised fine-tuning across 16 nodes with four GB300 GPUs each. It accepts conversations, input-output pairs or pre-tokenized examples. For conversation data, only assistant turns are supervised. Long samples are split into rows of 262,144 tokens.

The recipe starts from existing weights, resumes from saved steps and exports the result back to Hugging Face format. Its documentation distinguishes validated Blackwell configurations from unvalidated Hopper support. Export preserves tensor values and uses the original checkpoint’s configuration and tokenizer. The document identifies Apache 2.0 as the license for this implementation code.

Baidu’s published comparison uses the same evaluation framework, prompts and example counts across base checkpoints. HumanEval+ first-attempt success was 82.3%, against 69.5% for DeepSeek V4 Flash Base. On HELMET-Rerank, the scores were 56.2 and 60.0 respectively. These are company-reported comparisons of particular base models.

Commercial terms depend on related product and service revenue over the preceding 12 months. Within the permitted territory, revenue at or below $50 million qualifies for no model license charge, subject to attribution and a first-use notice. Above that threshold, separate prior written authorization is required.

The current Base model continues text through a completions endpoint. Baidu’s inference-plugin table lists the chat, reasoning and tool-calling version as forthcoming. The company says a post-trained ERNIE 5.5 Flash will follow, without specifying a release date.