Can You Run ERNIE 5.5 Flash Base Locally? 288 GB of FP8 and No Local Engine Yet

A stack of silver aluminium drive enclosures beside a monitor on a dark desk, lit with cyan-teal light

Not yet, though memory is not the blocker. Baidu released ERNIE 5.5 Flash Base as open weights on November 5, 2026, and its FP8 checkpoint takes about 288 GB, a little over half of what the largest Mac Studio holds. What is missing is software: neither llama.cpp nor mlx-lm can load its new ernie5_5_moe architecture today. This short guide gives the sourced numbers.

TL;DR: ERNIE 5.5 Flash Base is a 280B-parameter mixture-of-experts model with 14B active parameters, per the official model card. It is a base model, the checkpoint at the end of pre-training, not a chat or instruct model. Its FP8 weights take 288,416,689,496 bytes, per the safetensors index. llama.cpp and mlx-lm do not support it, and no GGUF or MLX build exists.

It is a base model, not a chat model

Baidu's card describes ERNIE-5.5-Flash-Base as "the checkpoint at the end of pre-training, before any mid- or post-training", meant as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. It is served through the completions endpoint, not a chat endpoint, and Baidu has not released an instruct or reasoning version alongside it. Do not expect it to follow instructions like an assistant.

What are ERNIE 5.5 Flash Base's specs?

Every figure below links to its source.

Spec	Value	Source
Total parameters	280B	model card
Activated parameters	14B per token	model card
Layers	43 (1 dense + 42 MoE)	model card, config.json
Experts	512 routed (top-10) + 1 shared per MoE layer	config.json
Attention	37 Heavily Compressed Attention (HCA) layers + 6 MLA layers with sparse attention (MLA-DSA)	model card
Pre-training tokens	33.5T	model card
Context length	131,072 tokens max in training, 1,048,576 for fine-tuning and inference	model card
FP8 weight size	288,416,689,496 bytes (about 288 GB)	FP8 model.safetensors.index.json
BF16 weight size	572,926,369,752 bytes (about 573 GB)	BF16 model.safetensors.index.json
Architecture	`ernie5_5_moe` (`Ernie5_5_MoeForCausalLM`)	config.json
Licence	ERNIE 5.5 Flash Base Model License Agreement v1.0	LICENSE

Sizes are the metadata.total_size field of each repository's safetensors index, in bytes. Baidu's own evaluation in the card puts it at 78.4 on MMLU-Pro (5-shot) against 69.2 for DeepSeek-V4-Flash-Base, with all models run through Baidu's pipeline; we have not reproduced it. More detail is in Baidu's technical blog.

How much memory does ERNIE 5.5 Flash Base need?

More than 288 GB for the FP8 weights alone, before context cache, runtime buffers and the operating system. The largest Mac Studio tops out at 512GB of unified memory, per Apple's Mac Studio tech specs, so the FP8 weights take about 56% of 512 GB counted in decimal (512,000,000,000 bytes) and about 52% of 512 GiB, our arithmetic. The BF16 release, at about 573 GB, is larger than that machine's memory. On memory alone the FP8 checkpoint fits a 512GB Mac Studio; on software it does not run there, because no Mac engine supports the architecture yet. We publish no speed figure, because no local engine can load the model to produce one.

What would a 4-bit build weigh?

No Q4 GGUF or MLX build exists on Hugging Face yet. As an estimate only: 280B parameters at 4.5 to 4.8 bits per weight, the range typical of llama.cpp's Q4_K_M, would come to about 158 to 168 GB (est., our arithmetic: parameters times bits per weight divided by 8), before any context cache. That is not a measured file size, and it assumes a converter that does not exist today.

Do llama.cpp and mlx-lm support ERNIE 5.5 Flash Base?

No, not as of November 5, 2026. Earlier ERNIE 4.5 models are supported, but ERNIE 5.5 Flash declares a new architecture with new attention layers.

llama.cpp: a search of the master branch converter, src and gguf-py finds no ernie5_5 or Ernie5_5 entry, and a GitHub search of its issues and pull requests returns none for ERNIE 5.5.
mlx-lm: mlx_lm/models has ernie4_5.py and ernie4_5_moe.py but no ernie5_5 file, and no issue or pull request mentions it.

Baidu's card lists server engines instead: SGLang, with a dedicated Docker image, and vLLM with the vLLM-ERNIE plugin, both shown on multi-GPU setups.

What does the licence allow?

It is not an OSI licence. Per the LICENSE file, non-commercial use such as research, evaluation and personal experimentation is free. Commercial use is free up to US$50 million of annual revenue from related products, with a one-time notice to Baidu and a "Powered by ERNIE" notice; above that it needs separate written authorization. The licence excludes the European Union, the United Kingdom and South Korea from its territory, and forbids using the model or its outputs to improve other AI models.

How does it compare with DeepSeek V4 Flash?

Baidu benchmarks it against DeepSeek-V4-Flash-Base, a model in the same size class. Our DeepSeek V4 Flash model page and DeepSeek V4.1 Flash memory guide cover a model you can already run locally. For a model that does not fit a Mac at all, see Kimi K3.1 Pro.

FAQ

Can a 512GB Mac Studio run ERNIE 5.5 Flash Base?

Not today. The FP8 weights, about 288 GB, use a little over half its memory, but mlx-lm and llama.cpp do not support the ernie5_5_moe architecture yet.

Is there a GGUF of ERNIE 5.5 Flash Base?

No. llama.cpp has no converter for the architecture, and Hugging Face lists no GGUF or MLX build.

Is ERNIE 5.5 Flash Base a chat model?

No. It is a pre-training checkpoint, served through the completions endpoint, with no instruct or reasoning release.

Sources

baidu/ERNIE-5.5-Flash-Base model card, config.json, model.safetensors.index.json and LICENSE; BF16 index; Baidu ERNIE blog, read November 5, 2026
llama.cpp and mlx-lm models directory, read November 5, 2026
Apple Mac Studio tech specs

Every number in this article is read from the source linked beside it or is our own arithmetic, labelled as such. How we work: about ModelFit. Corrections welcome: contact.