ERNIE 5.5 Flash Base
A lightweight base version large model in Baidu's ERNIE 5.5 series. By its naming, it is positioned for low-cost, low-latency deployment and high-concurrency scenarios, and belongs to the series' base versions.
Core Facts
Timeline (last 90 days)
Baidu released ERNIE 5.5 Flash Base, a sparse MoE foundation model with 280B total parameters, 14B activated parameters, and 33.5T tokens of training data.
33.5T tokens is at the higher end among current leading models, and the training process requires a stable data supply pipeline, deduplication and quality filtering mechanisms, and the ability to monitor and recover from training instability.
Comparison data published officially by Baidu shows that ERNIE 5.5 Flash Base scores 82.3 on the HumanEvo Plus coding test, while the comparison model DeepSeek V4 Flash Base scores 69.5.
ERNIE 5.5 Flash Base has 280 billion total parameters, with 14 billion parameters activated per token.
ERNIE 5.5 Flash Base was pre-trained on 33.5 trillion tokens and supports a maximum context of one million tokens.
Comparison data published officially by Baidu shows that ERNIE 5.5 Flash Base scores 78.4 on MMLU Pro, while DeepSeek V4 Flash Base scores 69.2.
Judging from the comparison charts released officially, ERNIE 5.5 Flash Base is benchmarked against DeepSeek V4.1 Flash, as well as previous-generation models such as its own ERNIE 3 Base version and MIMO 2.5 Base version.
Like the DeepSeek V4.1 Flash it is benchmarked against, ERNIE 5.5 Flash Base currently supports only plain text and does not yet support multimodal input such as images.
The ERNIE 5.5 Flash Base open-sourced this time is a Base model, that is, a version that has only completed basic training and has not yet undergone post-training such as preference alignment and reinforcement learning.
Baidu's ERNIE 5.5 Flash Base has been officially open-sourced, with 280B total parameters, 14B activated parameters, and support for million-level context.
7 more timeline events
All Facts (17)
The ERNIE 5.5 Flash Base open-sourced this time is a Base model, that is, a version that has only completed basic training and has not yet undergone post-training such as preference alignment and reinforcement learning.
65%Unverified33.5T tokens is at the higher end among current leading models, and the training process requires a stable data supply pipeline, deduplication and quality filtering mechanisms, and the ability to monitor and recover from training instability.
50%UnverifiedComparison data published officially by Baidu shows that ERNIE 5.5 Flash Base scores 78.4 on MMLU Pro, while DeepSeek V4 Flash Base scores 69.2.
50%UnverifiedComparison data published officially by Baidu shows that ERNIE 5.5 Flash Base scores 82.3 on the HumanEvo Plus coding test, while the comparison model DeepSeek V4 Flash Base scores 69.5.
50%UnverifiedERNIE 5.5 Flash Base has 280 billion total parameters, with 14 billion parameters activated per token.
50%UnverifiedERNIE 5.5 Flash Base was pre-trained on 33.5 trillion tokens and supports a maximum context of one million tokens.
50%UnverifiedERNIE 5.5 Flash Base has been open-sourced on Hugging Face, supports vLLM and SGLang inference deployment, and is offered in FP8 and BF16 formats, with no NVFP4 format available for now.
50%UnverifiedBaidu's official channels (such as the ERNIE Bot official website and the Qianfan platform) have not yet made the ERNIE 5.5 Flash Base model available, so there is a disconnect between open-sourcing and availability.
50%UnverifiedBaidu's ERNIE 5.5 Flash Base has been officially open-sourced, with 280B total parameters, 14B activated parameters, and support for million-level context.
50%UnverifiedJudging from the comparison charts released officially, ERNIE 5.5 Flash Base is benchmarked against DeepSeek V4.1 Flash, as well as previous-generation models such as its own ERNIE 3 Base version and MIMO 2.5 Base version.
50%UnverifiedLike the DeepSeek V4.1 Flash it is benchmarked against, ERNIE 5.5 Flash Base currently supports only plain text and does not yet support multimodal input such as images.
50%UnverifiedERNIE 5.5 Flash Base uses a 43-layer Transformer architecture with 64 attention heads and was trained on 33.5T tokens.
50%UnverifiedERNIE 5.5 Flash Base uses a hybrid MLA and HCA attention architecture, with the KV Cache portion optimized using MLA while most layers remain HCA structure.
50%UnverifiedThe Hacker News discussion of ERNIE 5.5 Flash Base had only 5 points and 2 comments.
50%UnverifiedThe name ERNIE 5.5 Flash Base suggests its positioning in lightweight design and inference efficiency.
50%UnverifiedBaidu released ERNIE 5.5 Flash Base, a sparse MoE foundation model with 280B total parameters, 14B activated parameters, and 33.5T tokens of training data.
50%UnverifiedBaidu was reported on Hacker News to have quietly launched the ERNIE 5.5 Flash Base model.
50%