ERNIE 5.5 Flash: Baidu's 280B Open MoE, Specs & Benchmarks
ERNIE 5.5 Flash Base is Baidu's open 280B MoE with 14B active, 33.5T tokens and 1M context. Specs, benchmarks vs DeepSeek V4 Flash, architecture and how to run it.
Quick answers
- What is ERNIE 5.5 Flash?
- ERNIE 5.5 is Baidu's efficient sparse MoE language model: 280B total parameters, 14B active per token, trained on 33.5T tokens. Baidu released the ERNIE 5.5 Base weights on November 5, 2026.
- Is ERNIE 5.5 Flash open source?
- Yes. The ERNIE-5.5-Flash-Base checkpoint is on Hugging Face in FP8 and BF16. The post-trained chat version of ERNIE 5.5 is coming next.
- How good is ERNIE 5.5 Flash?
- Baidu says the base model leads on 24 of 30 base-model benchmarks against DeepSeek V4 Flash Base, Hy3 Preview Base and MiMo V2.5 Base, with MMLU-Pro 78.4 vs 69.2 for DeepSeek.
- What context does ERNIE 5.5 Flash support?
- 128K tokens in pre-training and 1M tokens for fine-tuning and inference.
ERNIE 5.5 Flash is Baidu's new 280B open base model, and it beats DeepSeek V4 Flash Base on most tests. Baidu released ERNIE 5.5 Flash Base on November 5, 2026: raw pre-training weights with no instruction tuning, built so teams can run their own fine-tuning and RL on top.
This page covers ERNIE 5.5 specs, architecture, benchmarks, how to run it and what comes next.
ERNIE 5.5 Flash at a glance
Item | ERNIE 5.5 Flash Base |
|---|---|
Maker | Baidu ERNIE team |
Released | Nov 5, 2026 |
Size | 280B total, 14B active (MoE) |
Training data | 33.5T tokens, from scratch |
Context | 128K training, 1M fine-tuning and inference |
Layers | 43, hidden size 4096, 64 heads |
Experts | 512 routed (top-10) + 1 shared |
Weights | FP8 and BF16 on Hugging Face |
Serving | SGLang and vLLM reference configs |
ERNIE 5.5 Flash benchmarks
ERNIE 5.5 Flash Base | DeepSeek V4 Flash Base | Hy3 Preview Base | MiMo V2.5 Base | |
|---|---|---|---|---|
Total / active | 280B / 14B | 284B / 13B | 295B / 21B | 310B / 15B |
Context | 1M (inference) | 1M | 256K | 256K |
MMLU-Pro | 78.4 | 69.2 | — | — |
In Baidu's unified test pipeline, ERNIE 5.5 Base posts the best score on 24 of 30 base-model benchmarks across knowledge, reasoning, math, code, multilingual and long context. Its weakest spot is reranking: HELMET-Rerank 56.2 vs 60.0 for DeepSeek V4 Flash Base.
What's inside ERNIE 5.5 Flash
- Hybrid sparse attention: one layer in seven keeps full token-level access (MLA with sparse DSA); the rest use Heavily Compressed Attention, caching one entry per 128 tokens.
- Multi-Router MoE: several routers score experts together, so one router's mistake does not decide the pick.
- Latent experts: experts run in a 2048-d space, roughly halving traffic and expert FLOPs.
- mHC residuals and MultiMax head: four parallel residual streams and a learnable piecewise softmax temperature.
- ECHO: replaces the usual MTP module and doubles as a draft model for faster speculative decoding.
How to run ERNIE 5.5 Flash
Download baidu/ERNIE-5.5-Flash-Base from Hugging Face and serve it with SGLang or vLLM; Baidu's reference configs use 4-way parallelism on multi-GPU servers. This is a base model, so it continues text rather than following chat instructions. For chat or agent use, either fine-tune it or wait for the post-trained ERNIE 5.5 that Baidu says is next.
Turn base-model research into briefs with a Sumus AI Employee
Yes, you can keep up with every new base model without reading each paper. This post was written by a Sumus AI Employee: it read Baidu's release notes and model card and built the card above. A Sumus AI Employee can track releases, compare specs and post a short brief in your team chat.
What we're watching
- The post-trained ERNIE 5.5 chat model and its API price.
- Community fine-tunes on the base checkpoint.
- Independent long-context tests at 1M tokens.
FAQ
When was ERNIE 5.5 Flash released?
Baidu released the ERNIE 5.5 Base weights on November 5, 2026.
Is ERNIE 5.5 Flash a chat model?
Not yet. The release is a base checkpoint; the post-trained ERNIE 5.5 comes later.
How big is ERNIE 5.5 Flash?
280B total parameters with 14B active per token.
Where can I download ERNIE 5.5 Flash?
On Hugging Face as baidu/ERNIE-5.5-Flash-Base, in FP8 and BF16.
Are the ERNIE 5.5 Flash benchmarks independent?
They are Baidu's own, run in one pipeline against other base models.
Can a Sumus AI Employee track model releases like ERNIE 5.5 Flash?
Yes. A Sumus AI Employee reads launches, compares specs and reports in your team chat.
Update log
- — Nov 7: First published. ERNIE 5.5 Flash Base released Nov 5, 2026; post-trained model pending.