Baidu open-sources ERNIE 5.5 Flash Base on Hugging Face in FP8 and BF16
11/05/2026 — 11/06, 11:16·8 sources (all)·10 reports (all)
Key takeaways
- On Nov 5, Baidu open-sourced the ERNIE 5.5 Flash Base foundation model, a sparse MoE with 280B total parameters, 14B activated per token, pr…
Background & analysis
On November 5, 2026, Baidu open-sourced its foundation model ERNIE 5.5 Flash Base. The model uses a sparse MoE architecture with 280 billion (280B) total parameters and 14 billion (14B) active per token, and was pre-trained from scratch on 33.5 trillion (33.5T) tokens. Training context reaches up to 128K, while fine-tuning and inference support contexts of up to 1 million (1M) tokens.
According to Leiphone, Baidu claims ERNIE 5.5 Flash Base achieves SOTA among similarly sized models on benchmarks covering knowledge, reasoning, mathematics, coding, multilingual ability, and long context, outperforming DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base. QbitAI reported that the model sets new SOTA marks for its size class, though its summary gave no concrete scores, evaluation settings, or licensing details.
At 16:55 that day, Baidu's official X account (@Baidu_Inc) posted about ERNIE 5.5 Flash Base, describing it as an efficient sparse MoE language model trained on data spanning multiple domains and languages, and claiming SOTA on multiple benchmarks among similarly sized base models. The post pointed to two sources: the ERNIE blog (https://ernie.baidu.com/release/ernie-5-5-flash-base/) and Hugging Face (https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). The release page also appeared on Hacker News, where it reached the front page with 104 points and 57 comments.
On November 5, OSCHINA added that ERNIE 5.5 Flash Base carries no post-training at all and that Baidu emphasizes efficiency and cost-effectiveness. The Decoder followed on November 6, reporting that the model ships deliberately without fine-tuning or post-training, leaving it to developers to decide where to take it, with a post-trained version expected to follow. Weights are available on Hugging Face in both FP8 and BF16 formats. As of now, the public record consists mainly of model specifications and the release itself; none of the reports disclose the license, specific benchmark scores, or pricing.
AI-generated from 8 reports · updated 23 hours ago
Latest turnBaidu announced on X the release of ERNIE 5.5 Flash Base, an open-source foundation model featuring 280 billion parameters.
- Heat index
- 372
- All sources
- 8
- All reports
- 10
- First seen
- 1 day ago
Reports on this story headlines open the original
Baidu announced on X the release of ERNIE 5.5 Flash Base, an open-source foundation model featuring 280 billion parameters.
X：百度 (@Baidu_Inc)First-partyAI score 66Baidu has released ERNIE 5.5 Flash Base, a 280B sparse MoE base model with 14B active parameters, trained on 33.5T tokens prior to mid- or post-training stages.
MarkTechPostAI score 61Baidu has released ERNIE-5.5-Flash-Base, an open foundation model shipped deliberately without fine-tuning or post-training, leaving downstream direction to developers, with a post-trained version to follow, The Decoder reported. The MoE model has 280 billion total parameters with 14 billion active per input, was pre-trained on 33.5 trillion tokens with a 128,000-token training context, and its weights are on Hugging Face in FP8 and BF16.
The DecoderAI score 66
Baidu released ERNIE 5.5 Flash Base as an open-source foundation model on November 5. The MoE model has 280B total parameters with 14B activated per token, was pretrained on 33.5 trillion tokens with up to 128K context, and supports up to 1M tokens of context during fine-tuning and inference.
开源中国AI score 62Baidu has released ERNIE 5.5 Flash Base, with a release page on its official site drawing discussion on Hacker News. Public details so far cover only the model name and the release itself; capabilities and pricing are not disclosed in the available material.
Hacker News · 热榜AI score 52Baidu released and open-sourced ERNIE-5.5-Flash-Base, a sparse MoE model with 280B total parameters and 14B activated per token, pretrained from scratch on 33.5T tokens with a 128K training context and 1M-token context for fine-tuning and inference. Baidu's own benchmarks put it ahead of same-size base models such as DeepSeek-V4-Flash-Base on reasoning, coding, academic subjects and long-context tasks.
智东西AI score 62Baidu's official account announced ERNIE 5.5 Flash Base and linked to two places to explore it: the ERNIE blog and Hugging Face. The post itself gives no parameters or benchmark details, which are expected on those pages.
X：百度 (@Baidu_Inc)First-partyAI score 43Baidu released ERNIE 5.5 Flash Base, an efficient sparse MoE language model trained on data spanning multiple domains and languages. The company says it reaches SOTA on several benchmarks among similarly sized base models, and it is available to explore via the ERNIE blog and Hugging Face.
X：百度 (@Baidu_Inc)First-partyAI score 54Baidu open-sourced its latest base model, ERNIE 5.5 Flash Base, on November 5, according to QbitAI, which reports the model sets new SOTA results for its size class across multiple benchmarks. Specific scores, evaluation settings and license terms were not included in the report.
量子位AI score 54Baidu has open-sourced ERNIE 5.5 Flash Base, an MoE model with 280B total parameters and 14B activated per token, pretrained on 33.5 trillion tokens and supporting up to 1M-token context during fine-tuning and inference. Baidu says it reaches SOTA among same-size models across knowledge, reasoning, math, code, multilingual and long-context benchmarks, ahead of DeepSeek-V4-Flash-Base and others.
雷峰网AI score 65
Other stories people are talking about
- Heat index 61Hugging Face CEO Clément Delangue Discusses AI Risks and Open-Source Ecosystem1 sources
- Heat index 949OpenAI rolls out GPT-6 Cyber to Daybreak Red users with API, Codex access16 sources
- Heat index 791RisingClaude rolls out Personal Finance to US Pro, Max and Ultra users via read-only Plaid links13 sources
- Heat index 665SurgeAnthropic IPO valuation expectations fall to $1.2–1.3 trillion amid selloff9 sources
- Heat index 652SurgeThinking Machines Lab releases Inkling 2 Small amid reported $5B funding round9 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion