ERNIE 5.5 Flash Base: Weights, Memory and Deployment

ERNIE 5.5 Flash Preview is Baidu’s post-trained API model built on Base. Choose it for ready-to-prompt coding and agent tasks; keep this Base guide for adapting and serving the pre-training checkpoint. Preview has its own Qianfan prices and a separate OpenRouter free route.

This release is the checkpoint after pre-training and before mid-training or post-training. It is intended as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. Baidu now offers the separately post-trained ERNIE 5.5 Flash Preview through Qianfan. Choose that hosted assistant for ready-to-use chat; choose Base for adaptation and training. Preview’s announced future Apache 2.0 weight release does not change Base’s custom license. For scale, NVIDIA’s Megatron Bridge full-model SFT example uses 16 nodes with four GB300 GPUs per node; that is a documented training recipe, not a minimum for every fine-tune.

Baidu reports 280B total parameters, 14B activated per token and 33.5T pre-training tokens. The text-only model has 43 layers: one dense feed-forward layer and 42 MoE layers, each with 512 routed experts and top-10 routing. Its hybrid attention uses 37 Heavily Compressed Attention layers and six Multi-head Latent Attention layers with sparse attention.

The model card reports a 128K maximum training context and supports up to 1M tokens for fine-tuning and inference. A maximum context setting is not a promise that a particular machine can serve that many tokens at a useful rate; runtime allocation and concurrent requests still matter.

The official FP8 repository lists 288 GB of files split across 128 safetensors shards. The vLLM plugin estimates roughly 280 GB of FP8 weights and 560 GB of BF16 weights; those are weight estimates, while the repository’s actual size and the full serving footprint are separate questions. The 14B activated figure describes the parameters used for a token, not the amount of the expert pool that must be stored and made available to the server.

Treat 288 GB as the FP8 artifact size, not a GPU-memory minimum. Serving also needs room for runtime buffers, the KV cache, context length and other requests. Baidu’s card gives parallel serving recipes but does not specify a minimum GPU model or a complete peak-memory measurement. Choose hardware only after matching the exact checkpoint, engine and context to a tested configuration.

Baidu’s sample SGLang and vLLM launches use parallelism across devices and set a 131,072-token context. SGLang’s verified examples include two B300 or GB300 GPUs for FP8 and four B300 GPUs for BF16; these are tested recipes, not theoretical minimums. Its cookbook and the model card describe support up to 1,048,576 tokens, which raises the context and cache budget. Keep your chosen context and concurrency in the rental estimate instead of pricing the one-million-token headline by itself.

Baidu documents vLLM with its ERNIE plugin and SGLang with an ERNIE-specific image. The vLLM plugin supports this checkpoint on NVIDIA GPUs and pins vLLM 0.32.0. Its FP8 recipe uses the Humming MoE backend on Blackwell GPUs; on Hopper, Baidu says to omit that flag and keep vLLM’s default backend. The example sets 131,072 tokens as the model length. Use the SGLang cookbook for its matching image and launch flags.

The release uses a completion endpoint. Send a continuation prompt and inspect the generated text before adding an instruction wrapper or connecting the model to an agent. Base checkpoints are trained to continue text; they do not automatically follow chat instructions or emit the same structured tool calls as an instruction-tuned model.

A Baidu-affiliated commenter wrote in the model-card discussion that there was no official API service planned for this pre-training checkpoint, and directed developers to vLLM or SGLang. Hugging Face also showed no inference-provider deployment when checked on November 6, 2026. Treat this as the release-time access picture; the announcement publishes no per-token inference price.

Estimate a rented run from the full GPU allocation time and hourly price, then add storage and any data-transfer charge. Divide the session total by the number of outputs that meet your task’s quality bar. The FP8 artifact size and 14B active count do not provide a portable tokens-per-second or cost-per-task estimate; measure a representative prompt and context on the exact server you plan to use.

For a hosted trial of the post-trained model, OpenRouter lists baidu/ernie-5.5-flash-preview:free with free-tier rate limits.

The weights use Baidu’s ERNIE 5.5 Flash Base Model License Agreement v1.0, not Apache 2.0 or MIT. Its license grant excludes the European Union, the United Kingdom and South Korea. Outside those territories, non-commercial use is allowed subject to the agreement and acceptable-use policy.

The license measures related-product and service revenue over the rolling prior 12 months and aggregates entities under common control. At or below US$50 million, commercial use requires a one-time notice to Baidu and the “Powered by ERNIE” attribution; above that threshold, prior written authorization is required. The territory restriction also applies to outputs, and the same commercial conditions cover derivatives. Read the agreement for distribution, safeguards and acceptable-use terms before adapting or serving the weights.

Section 5.5 restricts using the licensed materials, Derivative Models or their outputs to improve other AI models or machine-learning systems, with an express exception for ERNIE 5.5 Flash Base itself and its Derivative Models. Read Section 1.5’s derivative definition when planning cross-model training: it includes models formed through distillation, knowledge transfer or training on outputs, subject to the agreement.