ERNIE 5.5 Flash Base
Baidu ERNIE 5.5 Flash Base
Baidu's open-weight 280B/14B hybrid-attention MoE base model released Nov 5, 2026: pre-trained on 33.5T tokens with 1M inference context, FP8/BF16 checkpoints; MMLU 93.3, HumanEval+ 82.3, RepoQA 91.7.
At a glance
- Context window
- 1M tokens
- Max output
- —
- Text input / output · per 1M tokens
- — / —
- Knowledge cutoff
- —
Benchmark snapshot
Best result in each categories
AcademicMMLU
93.3Normal (No Tools) · #2 of 114
CodingMBPP
81.5Normal (No Tools) · #17 of 67
Long ContextLongBench v2
48.7Normal (No Tools) · #19 of 22
ReasoningBBH
77.1Normal (No Tools) · limited comparability · #12 of 20
Safety & AlignmentSimpleQA
34.1Normal (No Tools) · #23 of 46
Benchmark details
9 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Academic | ||
| GPQA Normal (No Tools) · Scientific Reasoning | 60.2 | 6 / 15 |
| MMLU Normal (No Tools) · Knowledge Exams | 93.3 | 2 / 114 |
| MMLU-Pro Normal (No Tools) · Knowledge Exams | 78.4 | 69 / 153 |
| Coding | ||
| HumanEval Normal (No Tools) · Code Generation & Editing | 82.3 | 38 / 100 |
| LiveCodeBench Normal (No Tools) · Algorithmic Coding | 40.1 | 94 / 113 |
| MBPP Normal (No Tools) · Code Generation & Editing | 81.5 | 17 / 67 |
| Long Context | ||
| LongBench v2 Normal (No Tools) · Long Reasoning | 48.7 | 19 / 22 |
| Reasoning | ||
| BBH Normal (No Tools) · limited comparability · General Reasoning | 77.1 | 12 / 20 |
| Safety & Alignment | ||
| SimpleQA Normal (No Tools) · Honesty & Factuality | 34.1 | 23 / 46 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑Specifications
- Provider
- Baidu
- Model type
- Foundation model
- Released
- November 5, 2026
- Total parameters
- 280B
- Active parameters
- 14B
- Architecture
- Mixture of experts
- Context window
- 1M tokens
- Input
- Text
- Output
- Text
- Weights
- Open weights
- Licence
- Open weights (license unspecified)
- Commercial use
- Not documented
Capabilities
Model analysis
ERNIE-5.5-Flash-Base is Baidu's open-weight sparse Mixture-of-Experts (MoE) base model released on November 5, 2026, featuring 280B total parameters and 14B activated parameters per token, pre-trained from scratch on 33.5T tokens.
Architecture and Technical Highlights
- 280B-A14B Hybrid-Attention MoE: 43 layers (1 dense FFN + 42 Multi-Router Latent MoE) combining 37 Heavily Compressed Attention (HCA, 1 cache entry per 128 tokens, 512-dim heads) and 6 Multi-head Latent Attention layers with sparse attention (MLA + DSA, 256-dim heads); 512 routed experts (top-10) plus 1 shared expert per MoE layer in a 2048-d latent space, excluding 0.36B ECHO parameters.
- Context Window and Tokenizer: Pre-trained up to 128K (131,072 tokens) and supporting 1M (1,048,576 tokens) for fine-tuning and inference, with a 201,216-entry BPE vocabulary using right-to-left 3-digit number grouping.
- Key Innovations: Low-Rank Gated Attention and Virtual Head Attention (VHA, arXiv:2610.17959), Multi-Router Latent MoE (arXiv:2610.40577), ECHO backbone-reused multi-token prediction (arXiv:2610.24766), Manifold-constrained Hyper-Connections (mHC, expansion rate 4), and MultiMax output head, trained with distributed Muon and FlashMask-4.
Pre-training Benchmark Results
Under Baidu's unified evaluation pipeline across 30 benchmarks, ERNIE-5.5-Flash-Base leads comparably sized open base models: MMLU 93.3, MMLU-Redux 94.1, MMLU-Pro 78.4, CMMLU 90.0, GPQA 60.2, MGSM 87.0, MathBench-CN 80.8, HumanEval+ 82.3, MBPP+ 81.5, LiveCodeBench 40.1, RepoQA 91.7, and LongBench-v2 48.7.
Open Weights and Serving
FP8 (baidu/ERNIE-5.5-Flash-Base) and BF16 (baidu/ERNIE-5.5-Flash-Base-BF16) safetensors checkpoints are available on Hugging Face and ModelScope with SGLang and vLLM serving support under the ERNIE 5.5 Flash Base Model License Agreement v1.0.
Sources
Last reviewed 2026-11-16
Questions
- What are the parameter count and architecture of ERNIE-5.5-Flash-Base?
- ERNIE-5.5-Flash-Base has 280B total parameters and 14B activated parameters per token (excluding 0.36B ECHO parameters) across 43 layers, using 37 HCA layers, 6 MLA-DSA layers, and Multi-Router Latent MoE.
- What context length does ERNIE-5.5-Flash-Base support?
- It was pre-trained with a context length of up to 128K (131,072 tokens) and supports 1M (1,048,576 tokens) for fine-tuning and inference.
- Where are ERNIE-5.5-Flash-Base weights hosted and what license applies?
- FP8 and BF16 checkpoints are hosted on Hugging Face and ModelScope under the ERNIE 5.5 Flash Base Model License Agreement v1.0, with SGLang and vLLM serving support.
ERNIE 5.5 Flash Base vs competitors
Comparable results only · record is ERNIE 5.5 Flash Base's wins–losses
| Benchmark | ERNIE 5.5 Flash BaseThis model | vs Gemma 3 - 27B (IT)Record 6–2 | vs Llama3.1-405B InstructRecord 4–3 | vs Hunyuan-A13B-InstructRecord 3–3 | vs Mistral-Small-3.1-24B-Instruct-2503Record 5–1 |
|---|---|---|---|---|---|
| MMLU-Pro | 78.4 | 67.5 | 73.4 | 67.2 | 66.8 |
| MMLU | 93.3 | 78.6 | 88.6 | 88.2 | 80.6 |
| MBPP | 81.5 | 74.4 | 88.6 | 83.9 | 74.7 |
| LiveCodeBench | 40.1 | 29.7 | 30.2 | 63.9 | — |
| HumanEval | 82.3 | 87.8 | 89.0 | — | 88.4 |
| SimpleQA | 34.1 | 10.0 | 17.1 | — | 10.4 |
| BBH | 77.1 | 87.6 | 89.2 | 89.1 | — |
| GPQA | 60.2 | 36.8 | — | 49.1 | 44.4 |
Lineage
Predecessors and successors · oldest to newest
- ERNIE-4.5-300B-A47B-BaseJun 30, 2025
- ERNIE 5.5 Flash PreviewNov 11, 2026
Price & benchmark trends
Across curated predecessors and successors, ordered by release date. Points use listed rates and published results, not a history of price changes within one model. Related branches may appear; adjacent points do not imply direct succession. Undated models appear last.
API price across generations
At least two recorded values are needed for a trend. Available values appear below.
Input and output share one linear USD axis · uncached text rates · lower is cheaper. Retired models may retain their last recorded price. Rate tiers and context limits can differ; inspect each model's pricing conditions.
Orange marks this model. Hover, tap or focus the chart to see exact values. Missing values leave gaps. Horizontal spacing shows release order, not elapsed time.
| Model · released | Input / 1M | Output / 1M | Price effective |
|---|---|---|---|
| 1ERNIE-4.5-300B-A47B-BaseJun 30, 2025 | — | — | Not recorded |
| 2ERNIE 5.5 Flash BaseNov 5, 2026 · This model | — | — | Not recorded |
| 3ERNIE 5.5 Flash PreviewNov 11, 2026 | — | — | Not recorded |
Primary sources
- ERNIE 5.5 Flash Base technical reporternie.baidu.com
- GitHub repositorygithub.com
- Weights on Hugging Facehuggingface.co
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.
