Skip to main content

Novartis Says Two Tumour Cases in Children Are Linked to Its SMA Gene Therapy Zolgensma

Post a Comment

Baidu Open-Sources ERNIE 5.5 Flash: A 280-Billion-Parameter AI Model Anyone Can Download — With Limits

BEIJING — Baidu released the open weights of ERNIE 5.5 Flash Base on Thursday 5 November 2026, giving developers a free download of a 280-billion-parameter artificial intelligence model that uses only 14 billion parameters for each token it processes — a huge model built to run cheaply, but unfinished and with strings attached.

The weights are on Hugging Face in FP8 and BF16 formats, but this is the raw checkpoint at the end of pre-training, before any mid- or post-training. There is no chat, instruct or reasoning version yet, and the custom licence excludes the European Union, the United Kingdom and South Korea, and requires Baidu's written permission for commercial use above $50 million a year.

Illustration of a giant glowing neural-network diagram of many expert modules, with one small efficient core brightly lit inside the huge structure

Illustration: A giant mixture-of-experts network with only a small core lit up — just 14 billion of 280 billion parameters are active per token.

What Happened

On 5 November 2026, Baidu published the weights of ERNIE 5.5 Flash Base, its latest foundation model, on Hugging Face. The weights come in two formats, FP8 and BF16, and Baidu documents serving the model through SGLang and vLLM, using a completions endpoint rather than a chat interface.

The model is a sparse mixture-of-experts design, known as MoE. The network is divided into many specialist sub-networks, called experts, and only a small selection is switched on for each token of text. ERNIE 5.5 Flash Base has 280 billion parameters in total, but only 14 billion are active per token, about 5 percent — so each step costs roughly what a much smaller model would cost, while drawing on a far larger store of knowledge.

Baidu pretrained the model from scratch on 33.5 trillion tokens. Training used a context window of 128,000 tokens, and Baidu says the architecture extends to 1 million tokens for fine-tuning and inference. A million-token context would let a developer feed in an entire large codebase in a single pass.

What Baidu released is deliberately not a finished product. The checkpoint is the model exactly as it stood at the end of pre-training, before any mid-training or post-training — the stages in which a raw model is taught to follow instructions, answer conversationally, use tools, or reason step by step. Developers get a foundation to build on, not an assistant they can talk to. Teams that want a chatbot must do the fine-tuning themselves. Baidu says the post-trained ERNIE 5.5 Flash is coming soon, but has not given a date.

Baidu also published company-reported benchmark results, which should be read with that label attached. According to Baidu, the model scores 82.3 on HumanEval+, a test of code generation, against 69.5 for DeepSeek-V4-Flash-Base, and 40.1 on LiveCodeBench against 31.9, and it leads on 26 of 30 tests. Those figures have not been independently verified, and Baidu chose the rival and the tests.

The practical obstacles are substantial. The FP8 weights are about 288 gigabytes, so the files alone need a large machine or a multi-GPU server. The model's new architecture, ernie5_5_moe, is not yet supported by llama.cpp or mlx-lm, and no GGUF or MLX build exists yet. For now the model is a server workload, not something that runs on a laptop.

The licence is where 'open' meets its limits. ERNIE 5.5 Flash Base is governed by a custom document, the ERNIE 5.5 Flash Base Model License Agreement version 1.0, not a standard open-source licence. Its geographic grant excludes the European Union, the United Kingdom and South Korea. Commercial use is free of any model licence charge at or below $50 million a year in related revenue, subject to attribution and a one-time notice to Baidu; above $50 million, Baidu's separate written authorisation is required.

Background

Baidu's release is the latest move in an open-weight race that has defined the AI industry in 2026. In a few months, the largest Chinese AI laboratories have published a string of very large models with downloadable weights: Moonshot's Kimi K3.1 Pro, a 2.8-trillion-parameter model; Tencent's Hy4; and DeepSeek's V4 Flash Base, Baidu's main comparison point. Each release has pushed the same trade-off further — more total parameters, fewer active per token, longer context, cheaper inference.

The contest is as much about cost as capability. The 'Flash' in the name signals the trend: models built for speed and low cost per token, aimed at AI agents that call a model thousands of times a day. A model activating 14 billion parameters per token can, in principle, answer at close to the cost of a 14-billion-parameter model while drawing on 280 billion parameters of knowledge.

Baidu's turn to open weights is also a change of direction. Its chief executive had previously dismissed open-source software as a commercial dead end, and the company long kept its ERNIE models behind paid services. Baidu began revising that stance in June 2025, opening ERNIE 4.5 under the permissive Apache 2.0 licence. The new release goes further in scale but is tighter in its terms, and Baidu released no weights for ERNIE 5.0 or 5.1 in between.

This story is about Baidu's AI model business, not its transport business. Post #738 covered Baidu's Apollo World robotaxi simulation; the ERNIE 5.5 Flash Base release is a software release for developers, involving no vehicles or robotaxi fleet.

Illustration of developers downloading and opening a large AI model on a laptop, with code and weight files flowing from the screen

Illustration: Developers download the open weights of ERNIE 5.5 Flash Base and prepare to fine-tune the raw base model.

Nor is this the Tencent Hy4 story in posts #786 and #794. Hy4 is a rival model from a different company — same open-weight race, but a separate product with its own licence and benchmarks. The same applies to post #629, on Apple Intelligence in China and Alibaba's Qwen models, a partnership between two other companies.

The release also sits apart from other recent technology stories here: post #842 on METR and Anthropic's AI acceleration research, post #850 on OpenAI's disclosure of 948 Linux exploits, post #856 on the SpaceX and Qualcomm Starlink 5G announcement, post #864 on Google's $17-billion Wyoming data centre, and post #872 on OpenAI's Community Compact.

It is worth saying plainly what this release is not. Baidu is not the first Chinese company to open-source a model, and this article makes no such claim. DeepSeek, Alibaba, Tencent, Moonshot and Baidu itself have all published open models before. What is new is the combination: a 280-billion-parameter base model, 14 billion active per token, a million-token context, and a licence free for smaller users but closed to three major markets.

Why It Matters

The release matters first because of what it gives developers. Pre-training a model on 33.5 trillion tokens costs an enormous amount of computing power and money, beyond the reach of almost any university, startup or small company. By publishing the checkpoint, Baidu absorbs that cost and hands the result to anyone who can download and run it — for research, specialised models, or fine-tuning on private data.

It matters second because of the licence, which shows how the open-weight race is changing. The earliest wave of open models came with standard, permissive licences. This one draws lines on a map and on a balance sheet: developers in the European Union, the United Kingdom and South Korea cannot use it at all, and a company above $50 million in related revenue must negotiate with Baidu first.

It matters third because of the gap between a benchmark claim and a usable product. Baidu's reported lead — 82.3 against 69.5 on HumanEval+, 40.1 against 31.9 on LiveCodeBench, first on 26 of 30 tests — is the company's own numbers, on tests it selected, against a rival it chose. Independent verification has not happened. And even if the numbers hold, they describe a raw model, not a finished assistant.

Illustration of a world map with some regions dimmed and greyed out above a data-center skyline with glowing server lights

Illustration: Open access with limits — the licence excludes the European Union, the United Kingdom and South Korea.

It matters fourth because of the practical barriers. At about 288 gigabytes for the FP8 weights, with no support yet in llama.cpp or mlx-lm and no GGUF or MLX build, this is a model for organisations with multi-GPU servers and the expertise to run SGLang or vLLM. Downloading the weights is free, but using them is not cheap.

Finally, it matters because of what it signals about Baidu. A company whose chief executive once dismissed open source is now publishing its newest large model for anyone to download, sixteen months after its last open release. That suggests Baidu now believes a model's value lies less in keeping it private than in getting developers to build on it — and whether the bet pays off depends on whether the licence, rather than the model, makes people hesitate.

What Happens Next

The most immediate next step belongs to Baidu: the post-trained ERNIE 5.5 Flash, the chat and reasoning version, which the company says is coming soon. That will be the version most users can talk to, and it will show whether the strong base-model figures survive post-training. Until then, the only version in the world is the raw checkpoint.

The second step belongs to independent researchers. Baidu's figures — HumanEval+ 82.3, LiveCodeBench 40.1, a lead on 26 of 30 tests — await independent verification, and outside teams will now run their own evaluations, with different tests and methods. Those results will matter more than Baidu's own table.

The third step belongs to the tooling community. The new ernie5_5_moe architecture is not yet supported by llama.cpp or mlx-lm, and no GGUF or MLX build exists. Whether those tools add support, and how quickly, will decide how widely the model can be used outside well-funded server rooms.

The fourth step belongs to developers weighing the licence. Teams in the European Union, the United Kingdom and South Korea cannot use the model under its current terms, and companies above $50 million in related annual revenue must seek Baidu's written authorisation first. Smaller companies face a simpler choice — but still need the hardware and expertise to fine-tune a base model.

The wider race will not wait. Moonshot, Tencent, DeepSeek and Alibaba are all still publishing, and the next open-weight release may arrive within weeks. The question for Baidu is whether ERNIE 5.5 Flash Base becomes a foundation others build on, or another entry in a crowded leaderboard.

Sponsored: explore today's featured sponsored offers

FAQ

What is ERNIE 5.5 Flash Base?

ERNIE 5.5 Flash Base is a text-only, sparse mixture-of-experts AI model from Baidu, released on 5 November 2026. It has 280 billion parameters in total, with only 14 billion active per token, and was pretrained from scratch on 33.5 trillion tokens. It is a base checkpoint — the model at the end of pre-training — so it is a foundation to fine-tune, not a chatbot. The weights are on Hugging Face in FP8 and BF16 formats.

How big is the model, and can I run it on my own computer?

The FP8 weights are about 288 gigabytes, so the model needs a large machine or a multi-GPU server. It cannot yet run with llama.cpp or mlx-lm, which do not support its new ernie5_5_moe architecture, and no GGUF or MLX build exists. Baidu documents serving it through SGLang and vLLM. In practice it is a server workload, not a laptop model.

What do Baidu's benchmark results actually show?

According to Baidu, the model scores 82.3 on HumanEval+ against 69.5 for DeepSeek-V4-Flash-Base, and 40.1 on LiveCodeBench against 31.9, and leads on 26 of 30 tests. These are company-reported results that have not been independently verified; Baidu chose the tests and the rival. They are a claim to be checked, not a settled verdict.

What are the limits in the licence?

The model is governed by a custom licence, the ERNIE 5.5 Flash Base Model License Agreement version 1.0. Its geographic grant excludes the European Union, the United Kingdom and South Korea. Commercial use is free of any model licence charge at or below $50 million a year in related revenue, subject to attribution and a one-time notice; above $50 million, separate written authorisation from Baidu is required.

When will the finished ERNIE 5.5 Flash arrive?

Baidu says the post-trained ERNIE 5.5 Flash — the chat, instruct and reasoning version — is coming soon, but has not given a date. Until it arrives, the only version available is the raw base checkpoint, and anyone wanting a conversational assistant must fine-tune it themselves.

Sources

Labels:

Comments

Post a Comment

Popular posts from this blog

A CRISPR 'Gene Boost' Cut Seizures and Strengthened Muscles in Mice — Without Cutting DNA

SHANGHAI, Friday, November 6, 2026 — A compact new CRISPR system that turns helpful genes up instead of cutting DNA reduced seizures and prevented early deaths in mice modelling Dravet syndrome, and produced lasting improvements in muscle strength in mice modelling Duchenne muscular dystrophy, according to research published in Nature Biotechnology and reported on Monday, November 2, 2026. This article is for general information only and is not medical advice. The research described here was carried out in mice, and the treatment is not available for people. Anyone with questions about Dravet syndrome, Duchenne muscular dystrophy, or any medical condition should speak to a qualified clinician. Illustration: A CRISPR gene activation system targeting DNA in a laboratory, with molecular machinery and guide RNA strands approaching a gene's control region. What Happened Researchers led by Hainan Zhang of the Shanghai Institute of Materia Medica, Chinese Academy of Sciences, and co...
Post a Comment
Read more

CERN's LHCb Finds Evidence of One-in-Ten-Billion Particle Decay After 42-Year Search

GENEVA — Physicists working at CERN's Large Hadron Collider have found the first evidence from a single experiment of one of nature's rarest particle decays, a process so unlikely that the Standard Model of particle physics predicts it should happen only once in every ten billion beauty particles. The result, announced at a seminar at CERN on Tuesday, November 3, 2026, ends a search that began in the early 1980s with the CLEO experiment at Cornell University in the United States. It comes from the Large Hadron Collider beauty experiment, known as LHCb, which studies particles containing beauty quarks — heavier relatives of the quarks that make up much of the visible matter in the universe. The decay in question involves a neutral B0 meson, a short-lived particle made of a beauty quark and a down antiquark, transforming into a pair of muons, heavier cousins of the electron. According to the University of Birmingham, which plays a leading role in the LHCb collaboration, the mea...
Post a Comment
Read more

Oil Plunges Below $100 as Markets Rally After US Midterm Sweep

LONDON — Global markets delivered a dramatic split verdict on Thursday, 5 November 2026, as oil plunged below $100 a barrel while shares, gold and most major currencies rallied after Democrats swept the US midterm elections. Brent crude fell 97 cents, or 1.03 percent, to $93.65 a barrel as of 0211 GMT, while US West Texas Intermediate dropped $1.14, or 1.29 percent, to $86.98, according to Reuters. Both benchmarks had tumbled more than 8 percent in the previous session to their lowest since 31 August, and Brent is about 20 percent lower this week, heading for its biggest weekly drop since the pandemic shock of April 2020. Yet the S&P 500 closed 1.6 percent higher on Wednesday, European shares neared a two-month high, and gold hit its highest in more than two weeks. Investors are pricing in a lower chance of renewed US-Israeli strikes on Iran and a greater chance that the war which began at the end of February 2026 ends through negotiation. Illustration: An oil pumpjack beside ...
Post a Comment
Read more