Skip to content
Theorem Weekly TW markTWTheoremWeekly

The evidence behind the week in machine learning.

Model updates

Can ERNIE 5.5 Flash Base match DeepSeek V4 on open benchmarks?

Baidu releases a 280-billion-parameter MoE model without fine-tuning, inviting developers to test its raw pre-trained capabilities.

Key points

  • The model uses 280 billion parameters with 14 billion active per input.
  • Pre-training used 33.5 trillion tokens with up to 128,000 context.
  • Baidu claims it competes with DeepSeek V4 Flash and MiMo V2.5.
Can ERNIE 5.5 Flash Base match DeepSeek V4 on open benchmarks?
Depth
280 billion parametersThe model uses 280 billion parameters with 14 billion active per input · reported
14 billionThe model uses 280 billion parameters with 14 billion active per input · reported
33.5 trillion tokensPre-training used 33.5 trillion tokens with up to 128,000 context · reported
1 outlet · newest source 257h ago

The Decoder reported that Baidu has released ERNIE 5.5 Flash Base as an open-source foundation model. The release deliberately excludes fine-tuning or post-training steps.Baidu released a large AI model that developers must customize themselves before using.Baidu distributed ERNIE 5.5 Flash Base, a 280B MoE model, without alignment, emphasizing pre-trained benchmark competitiveness. Developers are expected to take the base weights and decide how to adapt the model for their specific use cases.

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

What was released or claimed

According to the report, the model employs a mixture-of-experts architecture with 280 billion total parameters. Only 14 billion parameters are active per input. Pre-training utilized 33.5 trillion tokens. The training context reached 128,000 tokens, though the design supports up to one million tokens during later stages.

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

Who it affects

The release targets machine learning developers and researchers. These groups can access the weights on Hugging Face. The availability in FP8 and BF16 formats suggests a focus on computational efficiency and flexibility for different hardware setups during the adaptation process.

What is still unknown

The performance of the model after community-led fine-tuning remains unverified by independent auditors. Baidu states that a post-trained version is supposed to follow, but no date or specific benchmark results for that version were provided in the report.

So what

The decision to ship a base model without post-training shifts the burden of alignment to the developer. This approach allows for maximum flexibility but requires significant computational resources from users to make the model useful for practical applications.

What the paper shows

Baidu’s claims focus on pre-trained benchmark performance rather than post-trained alignment. The company states the model competes with similarly sized open foundation models like DeepSeek V4 Flash and MiMo V2.5. These comparisons cover reasoning, code, and long-context tasks. The material does not provide specific scores, only the competitive positioning.

Evidence: reportedBaidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

aifleet-decoder

Why it matters

  • Researchers can evaluate raw pre-trained capabilities without proprietary alignment layers.
  • The 128,000-token training context limits immediate long-context testing despite future support.
  • FP8 and BF16 availability on Hugging Face enables efficient experimentation.
  • Comparison to DeepSeek V4 Flash provides a relevant baseline for open models.

Topics: open source, large language models, Baidu, mixture of experts, pre-training

Updated

Then / now / next

Then

Developers relied on fully aligned models or proprietary APIs for large-scale generation.

Now

Baidu offers a massive base model for custom fine-tuning by the community.

Next

Watch for the release of the post-trained version mentioned by Baidu.

Story timeline

  1. Then Developers relied on fully aligned models or proprietary APIs for large-scale generation.
  2. 05 Nov 2026 Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model
  3. Now Baidu offers a massive base model for custom fine-tuning by the community.
  4. Next Watch for the release of the post-trained version mentioned by Baidu.

The strongest doubt

Counterpoint from the sources

The model ships without fine-tuning, meaning it may lack safety or usability features present in post-trained versions.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

Does ERNIE 5.5 Flash Base support one million tokens now?

No, it was trained on 128,000 tokens but is designed to support one million tokens during fine-tuning and inference.

Is this model aligned for chat?

No, Baidu released it deliberately without fine-tuning or post-training, so developers must adapt it themselves.

Which models does it compete with?

Baidu claims it competes with DeepSeek V4 Flash and MiMo V2.5 on reasoning, code, and long-context tasks.

Connected stories

This storyHow does Phonon TTS achieve high quality on-device speech generation?How does Phonon TTS achieve highDoes AI-Generated Code Create Fragile SaaS Systems?Does AI-Generated Code Create FrCan LLMs Decode Unpublished Logic Notebooks Faster Than Human Scholars?Can LLMs Decode Unpublished LogiForum AI study finds mainstream chatbots misstate election facts nearly 40 percent of the timeForum AI study finds mainstream

Reading receipt

  • The model uses 280 billion parameters with 14 billion active per input.
  • Pre-training used 33.5 trillion tokens with up to 128,000 context.
  • Baidu claims it competes with DeepSeek V4 Flash and MiMo V2.5.

Still open: When will the post-trained version be released?

Still open: How does it perform on independent safety benchmarks?

Time spent: 0:00 · Depth: Full

Sources

  1. The Decoder

Related stories

Background reading