Baidu just open-sourced a model that beats DeepSeek
Open weights stopped being a goodwill gesture a while ago. They are how you take someone else's market while rivals sell subscriptions. On 5 November Baidu did exactly that: pushed ERNIE 5.5 Flash Base to Hugging Face — a base model anyone can fine-tune.
What's inside
Mixture of experts: 280B total parameters, 14B active per token. Pretraining on 33.5 trillion tokens. Context 128K during pretraining, up to 1M tokens for fine-tuning and inference. A million tokens in one request is enough to dump a pile of contracts and hunt for gaps in them.
Baidu claims SOTA among same-size models on knowledge, reasoning, math, code, multilingual work and long context. The list of models it beats includes DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base. These are direct rivals in the same weight class, not a company's own previous version.
Engineering vs marketing
The main design line is efficiency. Attention got rebuilt: Latent-Gated Sparse Attention, mixing sparse attention, low-rank gated attention and virtual head attention. Some layers compress history, others catch details. HCA and MLA work as a pair. Expert routing and training were reworked too.
The logic is simple: activating 14B of 280B parameters is cheaper than running the whole thing. On paper you have a 280B model; on the inference bill you pay for a slice.
Who should be uncomfortable
While Western labs debate whether to publish weights at all, Chinese ones ship them by the batch — with a benchmark report attached. DeepSeek just got a competitor in its own niche: a cheap base for fine-tuning. Closed API providers just got a client question: why am I paying per token when a base with million-token context and an open licence sits right there?
They have an answer: support, safety, SLA. That answer gets more expensive every month.
There is a second side. A benchmark is what the company chose to show. The comparison against DeepSeek-V4-Flash-Base and Hy3-preview-Base runs on a task set Baidu considers right. While the weights stay public, anyone can run their own tests and get a different number. That is the point of open weights — nobody owns the figure.
Sources: , Leiphone, weights on Hugging Face.
💬 What to argue about
- Does an open base with million-token context kill paid APIs for enterprise work — or will companies still pay for support and guarantees?
- Baidu benchmarks itself against DeepSeek and Hy3 on its own task set. Who should set the standard: the lab, an independent leaderboard, or the customer on their own traffic?
- If big open-weight models land every couple of months, who benefits from staying closed — the labs, or the companies building products on someone else's model?
Prediction: An independent run of ERNIE 5.5 Flash Base appears within a month, and the argument moves from benchmarks to real inference cost.
Disclaimer: not officially confirmed: no Baidu announcement in the source registry, data comes from industry outlets
🤖 AI summary
- Date: 2026-11-05
- Source: www.qbitai.com
- Tags: Baidu, ERNIE, OpenSource
- Key facts:
- Open weights stopped being a goodwill gesture a while ago.
- They are how you take someone else's market while rivals sell subscriptions.
- On 5 November Baidu did exactly that: pushed ERNIE 5.5 Flash Base to Hugging Face — a base model anyone can fine-tune.
- Mixture of experts: 280B total parameters, 14B active per token.
- Pretraining on 33.5 trillion tokens.
- Context 128K during pretraining, up to 1M tokens for fine-tuning and inference.
- A million tokens in one request is enough to dump a pile of contracts and hunt for gaps in them.
- Baidu claims SOTA among same-size models on knowledge, reasoning, math, code, multilingual work and long context.
Source: www.qbitai.com · post in Telegram