SSyncDev

ERNIE 5.5 Flash Base

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

ERNIE 5.5 Flash Base is Baidu's open 280B-A14B text MoE pre-training checkpoint, with a 1M-token window, FP8 and BF16 weights for fine-tuning.

Key specifications

Specification score
31.6
Context window
1M
License
ERNIE 5.5 Flash Base Model License Agreement v1.0 (custom; secondary sources report EU/UK/South Korea exclusions and a revenue threshold)

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is ERNIE 5.5 Flash Base?

This is raw material, not a chatbot. Baidu released the checkpoint from the end of pre-training, before any mid-training, instruction tuning or RLHF, and it has no chat template. It is meant for teams who want to fine-tune or continue training on their own data.

Under the hood it is a text-only sparse mixture-of-experts with 280B total parameters and 14B active per token.

Architecture in brief

Item Detail
Hugging Face repo baidu/ERNIE-5.5-Flash-Base (FP8), baidu/ERNIE-5.5-Flash-Base-BF16
Parameters 280B total / 14B active; metadata shows 286B including ECHO
Layers 43: 1 dense FFN plus 42 MoE
Experts 512 routed (top-10) plus 1 shared per MoE layer
Attention 37 heavily compressed attention layers plus 6 MLA-DSA layers
Context Trained at 128K; config enables 1M for fine-tuning and inference
Training 33.5T tokens, Muon optimiser, FP8 mixed precision
Vocabulary 201,216 tokens

The 286B-versus-280B gap is explained by ECHO, which replaces a separate multi-token-prediction module with the final layer and powers speculative decoding. Its parameters show up in the weight files but not in the headline count.

Serving it

Both reference recipes use four-way parallelism. With SGLang, there is a prebuilt lmsysorg/sglang:dev-ernie-55-flash image; with vLLM you need version 0.32.0 plus Baidu's vLLM-ERNIE plugin. Being a base model, it is served through the /v1/completions endpoint, not chat completions.

Memory is tight. FP8 weights are about 280 GB and BF16 about 560 GB, so four 80 GB cards leave roughly 40 GB for KV cache and runtime with FP8. Long contexts or real batch sizes will need more GPUs or lower concurrency.

The licence is not Apache

Unlike ERNIE 4.5, this checkpoint ships under the ERNIE 5.5 Flash Base Model License Agreement v1.0. Reporting describes exclusions for the European Union, the United Kingdom and South Korea, and a revenue condition for very large commercial users. The threshold is not stated in what we read. Read the LICENSE file in the repository before any commercial use; it is the only authoritative text.

Should you start from it?

Publisher-reported results put it ahead of comparable base models on most standard sets (93.3 on MMLU 5-shot, 78.4 on MMLU-Pro, 82.3 on HumanEval+), though it trails on a few, including ChineseSimpleQA and InfiniteBench. For a fair decision, fine-tune it and one alternative on the same small slice of your data and compare held-out loss and task accuracy.

If you want a tuned model instead, ERNIE 5.5 Flash Preview is the first post-trained checkpoint built on this base. Earlier open ERNIE weights, such as the Apache-licensed ERNIE 4.5 releases, sit on the Baidu Hugging Face page, and the text-generation directory lists other base models.

How to evaluate ERNIE 5.5 Flash Base

Compare the model on the work you need to complete.

Tasks and inputs

ERNIE 5.5 Flash Base is listed for Language & text, Text generation, Coding & Code Generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 1M context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under ERNIE 5.5 Flash Base Model License Agreement v1.0 (custom; secondary sources report EU/UK/South Korea exclusions and a revenue threshold). Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

No. It is the end-of-pre-training checkpoint with no instruction tuning, RLHF or chat template, served via the completions endpoint.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Tencent

Hy4 Flash

Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.

1.05M context

Alibaba

Qwen4 Max

Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.

1M context

Sarvam

Sarvam-105B-1711

Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.

— context

Alibaba

Qwen4 Max Preview

Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.

1M context

Google

Gemini 4 Carbon

Gemini 4 Carbon is Google's strongest Gemini 4 reasoning model for long-horizon agentic coding, scientific research and cyber defence, with a 1M context.

1.05M context

Anthropic

Claude Fable 5.2

Claude Fable 5.2 is Anthropic's top model for demanding reasoning and long-horizon agentic work, with a 1M context at $8/$40, 20% below Fable 5.1.

1M context