Baidu's ERNIE 5.5 Flash Packs 280B Parameters Into a 14B Compute Path

Baidu has announced ERNIE 5.5 Flash Preview, the post-trained version of a 280-billion-parameter sparse foundation model. Its untuned base checkpoint is already available without instruction tuning, reinforcement learning from human feedback, or a chat template. Baidu says the Preview weights will follow by November 25.

For developers, the model’s appeal comes from its compute profile: only 14 billion parameters activate for each token, despite all 280 billion parameters remaining in memory. Baidu reports generation above 200 tokens per second, support for contexts up to one million tokens, and competitive coding and knowledge scores. Those claims come with substantial hardware requirements, limited independent testing, and a custom license that restricts deployment in several regions.

A 280B model with a 14B compute path

The mixture-of-experts design stores a large pool of specialized feed-forward networks and routes each token through a small subset. This reduces arithmetic per token while preserving the capacity of the larger model. The full parameter set must still reside in memory, so sparse activation lowers compute costs more than storage requirements.

Baidu’s hybrid attention stack compresses the key-value cache used to retain earlier context during generation. That compression helps make long contexts practical, although memory use still rises with sequence length, batch size, and concurrency.

Baidu attributes part of the reported decoding speed to ECHO, a speculative decoding method that reuses the backbone’s final layer to draft future tokens. The full model verifies those drafts, avoiding the memory and operational cost of serving a separate draft model.

Thirty-three trillion training tokens

Pre-training consumed 33.5 trillion tokens using the Muon optimizer and FP8 mixed precision. Baidu gradually extended the training context to 128,000 tokens, while post-training and inference support a stated maximum of one million tokens.

The training corpus represents roughly 120 tokens for every total parameter. That ratio is useful for comparing data scale, though it does not describe the compute applied to each token because the model activates only a fraction of its parameters during a forward pass.

The one-million-token limit is a supported context ceiling rather than evidence of consistent recall across the entire window. Long-context accuracy, latency, and memory consumption still require testing against the intended workload.

Benchmark results split by task

Baidu reports that the base checkpoint beats DeepSeek-V4-Flash-Base on 26 of 30 tests in its evaluation pipeline. The largest published gains appear in broad knowledge evaluation and single-pass code generation.

Different evaluation harnesses, prompts, scaffolds, and post-training configurations limit direct comparison between Baidu’s base-model results and independent agent runs. The available evidence supports strong knowledge recall and code completion, while sustained tool use and environment interaction remain weaker areas.

Artificial Analysis has not published an ERNIE 5.5 evaluation, leaving latency, quality, and cost without a widely used third-party reference point.

Memory is the deployment constraint

Baidu provides four-way parallel configurations for SGLang and vLLM. A four-GPU node with 80 GB cards offers 320 GB of raw device memory, leaving roughly 40 GB after FP8 weights alone. The inference engine, communication buffers, key-value cache, and concurrent requests consume the remaining capacity.

Long contexts and production batch sizes may require more GPUs, lower-precision quantization, reduced concurrency, or cache offloading. BF16 deployment starts at about 560 GB for weights and needs further capacity for the serving runtime.

The reported speed above 200 tokens per second should be treated as a vendor result until reproduced on a specified GPU configuration, batch size, context length, and serving stack. Throughput and per-request latency can diverge sharply under continuous batching.

The Preview announcement includes no hosted API, service-level agreement, or public price sheet. Current adoption therefore requires self-hosting and direct management of capacity, observability, upgrades, and licensing compliance.

The license narrows the map

ERNIE 5.5 Flash uses a custom license that lists the European Union, United Kingdom, and South Korea among its excluded regions. It also includes a revenue-based condition for very large commercial users.

Teams serving customers across multiple jurisdictions need to check the license text against their legal entity, deployment location, user location, revenue, redistribution plans, and derivative models. Apache 2.0 assumptions from other open-weight releases do not apply to this checkpoint.

A crowded 280B class

Baidu evaluated these models through the same internal pipeline. Their similar shapes show a recurring design choice among Chinese AI labs: hundreds of billions of stored parameters paired with roughly 13 billion to 21 billion active parameters per token.

ERNIE 5.5 Flash differentiates itself through hybrid compressed attention, ECHO speculative decoding, and Baidu’s throughput claim. Each feature targets serving efficiency, which becomes decisive when inference volume outweighs experimentation costs.

Deployment fit by workload

Teams with sufficient GPU capacity, compatible deployment geography, and a need for text generation or custom post-training have the clearest reason to evaluate ERNIE 5.5 Flash. Production testing should measure task quality, time to first token, sustained throughput, cache growth, and agent reliability under the same serving conditions planned for release.