Model library / baidu
ERNIE-5.5-Flash-Base is a 286 billion-parameter sparse Mixture-of-Experts language model released by Baidu.
10 of 512 experts active per token
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
ERNIE-5.5-Flash-Base is an efficient sparse MoE model designed for general-purpose language tasks. Baidu describes it as a pre-training checkpoint suitable for domain adaptation, instruction tuning and reinforcement-learning research, offering strong performance on knowledge, reasoning, code and long-context benchmarks. The architecture uses 512 experts with ten active per token and hybrid attention layers.
In practice the model contains 286.2 billion parameters, stores weights in fp8 format and supports a context window of up to 1,048,576 tokens. It is released under an "other" licence, requiring users to check the specific terms. Deployments must handle the large memory footprint and the MoE routing overhead, while benefiting from the very long context capability.
Specification
| Repository | baidu/ERNIE-5.5-Flash-Base |
| Publisher | baidu |
| Published | 21 October 2026 |
| Architecture | Ernie5_5_MoeForCausalLM |
| Model type | ernie5_5_moe |
| Parameters | 286,245,697,882 |
| Experts | 512 total, 10 active per token |
| Layers | 43 |
| Hidden size | 4,096 |
| Attention heads | 64, 1 key/value heads |
| Context length | 1,048,576 tokens |
| Vocabulary | 201,216 tokens |
| Quantisation | fp8 |
| Licence | other |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 533.2 GiB | Full precision as released |
| fp8 | 266.6 GiB | 8-bit, near-lossless on most models |
| int4 | 133.3 GiB | 4-bit, smallest footprint |
The cache costs 86 KiB per token at 16-bit, so the full 1,048,576-token context of one request needs about 86.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses compressed key-value attention, so the real cache is a fraction of the figure above.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | more than 8 | — | — |
| RTX 5090 32GB | more than 8 | — | — |
| L40S 48GB | 8 | 57.1 GiB | about 695,661 tokens |
| A100 80GB | 4 | 9.5 GiB | about 115,286 tokens |
| H100 80GB | 4 | 9.5 GiB | about 115,286 tokens |
| RTX PRO 6000 96GB | 4 | 67.1 GiB | about 817,588 tokens |
| H200 141GB | 4 | 229.1 GiB | the full 1,048,576 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve baidu/ERNIE-5.5-Flash-Base \ --tensor-parallel-size 8 \ --max-model-len 692224 \ --gpu-memory-utilization 0.90
Questions
The weights take about 269 GiB at the released precision. No single card in the table above holds that, so it needs several GPUs or a lower precision.
The repository declares other. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 1,048,576 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
Ling-3.1-flash-1019 is a 572.1 billion-parameter causal language model from inclusionAI, designed for advanced reasoning and agen…
Ling-3.1-flash-base is a 560 billion-parameter hybrid MoE base checkpoint from inclusionAI, provided in bfloat16 for further trai…
Ling-3.1-flash is a 572.1 billion-parameter hybrid reasoning language model with a 262,144-token context window.
Beam is an open-source, instruction-tuned language model from Reflection using a sparse mixture-of-experts architecture.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.