Primary use stated or inferred cautiously from official documentation.
ERNIE 5.5 Flash Base
Baidu base checkpoint released on November 5 with 280 billion total parameters and 14 billion active per token. It was trained on 33.5 trillion tokens and takes and generates text. The FP8 and BF16 repositories are two precisions of the same pretrained release.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | baidu/ERNIE-5.5-Flash-Base |
|---|---|
| Model type | Text and reasoning · Code |
| Access model | Restricted weights |
| License | ERNIE 5.5 Flash Base Model License Agreement v1.0 |
| Deployment | Self-hosted |
| Release | 05 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text |
| Exit | Text |
| Context window | 128K during pretraining; 1M for tuning and inference |
| maximum output | Not published |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · ModelScope · Transformers |
Domain adaptation and post-training research with fixed settings and precision. The provider states 128K pretraining context and up to 1M for fine-tuning and inference. The base checkpoint is not tuned as a conversational assistant.
The license excludes the European Union, United Kingdom and South Korea from its territory. It also sets commercial-use and distribution conditions. Visible files do not grant permission in excluded territories. Performance of a future post-trained version is not attributed to this checkpoint.
What it can do
Orientation
Domain adaptation and post-training research with fixed settings and precision. The provider states 128K pretraining context and up to 1M for fine-tuning and inference. The base checkpoint is not tuned as a conversational assistant.
Context
128K during pretraining; 1M for tuning and inference · Not published
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No API price has been verified for the base checkpoint. Local compute and license compliance involve separate costs. Check the official commercial-use requirements; this is not an Apache license or a guarantee of free inference.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
Baidu reports 93.3 on MMLU with five examples, 82.3 on HumanEval+ pass@1 with one example and 48.7 on LongBench-v2 exact match without examples. The provider uses a common protocol for base checkpoints. Inferama has not reproduced these figures or verified separate scores by precision.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| MMLU (base, 5-shot EM) | 93,3 / 100 | Score | View source ↗ |
| HumanEval+ (base, 1-shot pass@1) | 82,3 / 100 | Score | View source ↗ |
| LongBench-v2 (base, 0-shot EM) | 48,7 / 100 | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
ERNIE 5.5 Flash Base
Version added to Inferama's verified catalog.