--- title: Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B... description: Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a sparse mixture-of-experts model with 280B total parameters and 14B active per token.... canonical: https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B active parameters, with day-0 vLLM support | daily.dev og:description: Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a sparse mixture-of-experts model with 280B total parameters and 14B active per token.... og:url: https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq og:image: https://api.daily.dev/og/posts/5P08CWsVq.png og:image:alt: Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B active parameters, with day-0 vLLM support og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B active parameters, with day-0 vLLM support **[Collections](https://daily.dev/sources/collections)** · 2 min read · 11 upvotes · 1 comments ## Summary Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a sparse mixture-of-experts model with 280B total parameters and 14B active per token. It was pretrained from scratch on 33.5T tokens, with a 128K training context extendable to 1M tokens. vLLM supported it from day zero, and runs go through Baidu's vLLM-ERNIE plugin on vLLM 0.32.0 in FP8 or BF16. On HumanEval+ it scores 82.3% versus 69.5% for DeepSeek-V4-Flash-Base, and on LiveCodeBench pass@1 it scores 40.1% versus 31.9%. Weights are on Hugging Face and ModelScope. ## Content Baidu has released ERNIE-5.5-Flash-Base, the raw pre-training checkpoint of its new sparse Mixture-of-Experts model. It has 280B total parameters, with 14B active per token. vLLM announced day-0 support. ## Training and context The model was pretrained from scratch on 33.5T tokens with a training context of 128K. For fine-tuning and inference, that can be extended to 1M tokens. Baidu is positioning the open base weights for domain adaptation, instruction tuning and RL research. ## Architecture - Hybrid attention: 37 compressed-attention layers (Heavily Compressed Attention) plus 6 layers that combine Multi-head Latent Attention with DSA. - A Multi-Router Latent MoE with 512 routed experts and a shared expert. - mHC residual streams and a MultiMax output head. - ECHO, which replaces multi-token prediction. It reuses the final layer as a speculative draft for decoding. ## Benchmarks On Baidu's own benchmark pipeline, the model scores best on 24 of 30 base-model tests. Its comparisons are against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, and it does well on most exam, reasoning and code benchmarks. Some of the numbers against DeepSeek-V4-Flash-Base: | Benchmark | ERNIE 5.5 Flash Base | DeepSeek-V4-Flash-Base | |---|---|---| | MMLU-Pro | 78.4 | 69.2 | | HumanEval+ (1-shot) | 82.3% | 69.5% | | LiveCodeBench pass@1 (1-shot) | 40.1% | 31.9% | These are Baidu's evaluations, run under its own setup, so independent replication will matter. ## Serving Weights are on Hugging Face and ModelScope in FP8 and BF16. To serve them, use Baidu's vLLM-ERNIE plugin with vLLM 0.32.0. ECHO runs through vLLM's MTP path. The recipes cover expert parallelism, KV offload and prefill/decode disaggregation, and runs have been verified on NVIDIA GB200 and H100. ## License The license is free for organizations under $50M in revenue. It excludes the EU, the UK and South Korea, which will rule the model out for some teams regardless of how good the numbers look. ## Questions this post answers ### How many parameters does Baidu's ERNIE-5.5-Flash-Base model activate per token? ERNIE-5.5-Flash-Base has 280 billion total parameters but only activates 14 billion per token, since it uses a sparse mixture-of-experts architecture with 512 routed experts plus a shared expert. It was pretrained from scratch on 33.5 trillion tokens with a 128K token training context, extendable to 1M tokens for fine-tuning and inference. _Comparing MoE model specs before picking one to fine-tune gets easier when tracked on daily.dev._ ### How does ERNIE-5.5-Flash-Base compare to DeepSeek-V4-Flash-Base on coding benchmarks? ERNIE-5.5-Flash-Base outperforms DeepSeek-V4-Flash-Base on code benchmarks under a 1-shot evaluation setup. It scores 82.3% on HumanEval+ versus 69.5% for DeepSeek-V4-Flash-Base, and 40.1% versus 31.9% on LiveCodeBench pass@1. _Developers weighing base models for code generation can follow benchmark comparisons like this on daily.dev._ ### Does vLLM support Baidu's ERNIE-5.5-Flash-Base model out of the box? Yes, vLLM supported ERNIE-5.5-Flash-Base from day zero through Baidu's vLLM-ERNIE plugin running on vLLM 0.32.0. The model runs in FP8 or BF16, with recipes covering expert parallelism, KV offload, and prefill/decode disaggregation, verified on NVIDIA GB200 and H100 hardware. _Teams deploying new MoE checkpoints can track day-0 serving support like this through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 93 comments across hackernews (as of 2026-11-06). **TL;DR:** Discussion is mostly tangential jokes about the ERNIE name and Bert/Ernie references, but the substantive threads focus on the value of releasing a base (non-instruction-tuned) model, its size/hardware fit, and confusion over the license's exclusion of the EU, UK, and South Korea. **Sentiment:** 25% positive · 45% mixed · 30% skeptical **The case for** - Several commenters welcomed the release of a raw base model, arguing it's more useful than instruction-tuned releases for those who want to fine-tune or train in-house behaviors. - Some noted it appears more capable than comparably sized base models like Deepseek-v4-Flash-Base given its smaller active parameter count. - It was pointed out that the model's size makes it feasible to run on dual 128GB setups at decent quantization. **The pushback** - Many found the license's blanket exclusion of the EU, UK, and South Korea confusing or hypocritical, seeing it as a way to offload legal/regulatory risk onto users rather than a genuine restriction. - Some questioned whether comparisons to Deepseek were fair since different Deepseek versions/architectures were being referenced. - One commenter dismissed it outright as low-quality, calling it "slop." **By community** - hackernews (mixed): A long tangent of jokes about the ERNIE name dominates, while the more substantive sub-threads are split between enthusiasm for an open base-weights release and confusion/criticism over the EU/UK/South Korea license carve-out. **Hottest debate:** Whether the license's exclusion of the EU, UK, and South Korea is a meaningful restriction or just a liability-shifting formality that won't actually be enforced. **Open questions** - Has any court actually ruled that model weights are copyrightable, which would affect how the licensing restriction could be enforced? - What exactly drove Baidu to exclude those specific regions — direct AI Act compliance costs, competitive reasons, or something else? - Will the model recreate the strong early reception ERNIE 4 got before Deepseek and GLM overtook it? **Highlights** > The license seems really restrictive for something intended to be built on. Entire countries are forbidden to use it at all (European Union, the United Kingdom and South Korea), and you are responsible for making sure no one in those places uses your derivative. > — [mkl on hackernews · 7 comments](https://news.ycombinator.com/item?id=50378719) > That's where you've misunderstood Baidu's intent. They don't care to enforce it, only to check the box that says they told you open weight models are all but illegal in EU, UK, and South Korea. "You cannot use our models in these regions." Now it's on the EU, UK and South Korea to stop users from downloading and running those models. All that said, the EU, UK and South Korea, if they want to, are all more than capable of making life a living perdition for people who do run those models. So I wouldn't say it's "not enforceable". Only that it probably won't be enforced in the fashion one would think these rules would be enforced. > — [bilbo0s on hackernews](https://news.ycombinator.com/item?id=50379890) > Many others, e.g. Google and NVIDIA, have also released such BASE models. For those with the knowledge and resources needed to finalize the training, BASE models are much more useful than completely trained models. BASE models are the next best thing after truly open-source LLMs, where you would have access to the complete dataset used for training. Besides the fact that nobody who had used a really big dataset wants to show it, due to possible legal problems, having access to the original dataset would not be helpful for most people, because they would not be able to use it to make an LLM, due to its huge size and due to the vast resources required for pre-training. On the other hand, continuing training from a BASE model, which is greatly compressed in comparison with the original dataset, is feasible for many. > — [adrian\_b on hackernews](https://news.ycombinator.com/item?id=50378010) > Mostly yes, you are wrong. Active parameter count really drives inference speed not ram requirements. For this model, assuming you want the full 1mm context length at 4-bit quants, you need something like 175 GB total VRAM. ~150GB of that will be weights (varies a bit by quant format) then ~3-5GB for kv cache (might be lower here, I think they're using mostly HCA compressed cache) and then I'll allocate a couple gb for overhead/concurrency (say another ~10GB as a safety buffer). --- So the good news is that it comfortably fits into 2x128 at 4-bit, and we can realistically stretch to a 6-bit quant and make it just fit (~244GB needed for the full 1mm context). > — [horsawlarway on hackernews](https://news.ycombinator.com/item?id=50379791) > Absolutely glad to see the release of a base model, I think the trend away from that has been an issue for the field. > — [karahime on hackernews](https://news.ycombinator.com/item?id=50378100) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50377078) · 167 points · 93 comments ## Community discussion Top comments from developers on daily.dev. **@skepagent** · 0 upvotes > Base 权重加 128K 上下文这个组合很实在，拿来做领域微调和 RL 研究比直接给个 instruct 版更灵活。比较好奇的一点是 ECHO 用最后一层当 speculative draft，实测加速比和直接上 MTP 相比怎么样，文章里只讲了机制，没给速度对比。 ## Similar posts on daily.dev - [Baidu launches new generation of Ernie AI](https://daily.dev/posts/baidu-launches-new-generation-of-ernie-ai-w1y8hd0zv) · InfoWorld · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#vllm](https://daily.dev/tags/vllm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B active parameters, with day-0 vLLM support","url":"https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq"},"datePublished":"2026-11-05T14:15:37.353Z","dateModified":"2026-11-06T03:49:21.367Z","description":"Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a sparse mixture-of-experts model with 280B total parameters and 14B active per token....","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":11},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"llm,vllm,mixture-of-experts","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Baidu releases ERNIE 5.5 Flash Base, a 280B MoE with 14B active parameters, with day-0 vLLM support"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq","comment":[{"@type":"Comment","text":"Base 权重加 128K 上下文这个组合很实在，拿来做领域微调和 RL 研究比直接给个 instruct 版更灵活。比较好奇的一点是 ECHO 用最后一层当 speculative draft，实测加速比和直接上 MTP 相比怎么样，文章里只讲了机制，没给速度对比。","datePublished":"2026-11-06T02:50:39.355Z","url":"https://daily.dev/posts/5P08CWsVq#c-8yacqldYZ","author":{"@type":"Person","name":"kris gu","url":"https://daily.dev/skepagent","image":"https://lh3.googleusercontent.com/a/ACg8ocJnnwFKNx7PJNhb7W6mFWr2k2SV_VHW_9ainY71DX_dv79frQ=s96-c"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/baidu-releases-ernie-5-5-flash-base-a-280b-moe-with-14b-active-parameters-with-day-0-vllm-support-5p08cwsvq#faq","mainEntity":[{"@type":"Question","name":"How many parameters does Baidu's ERNIE-5.5-Flash-Base model activate per token?","acceptedAnswer":{"@type":"Answer","text":"ERNIE-5.5-Flash-Base has 280 billion total parameters but only activates 14 billion per token, since it uses a sparse mixture-of-experts architecture with 512 routed experts plus a shared expert. It was pretrained from scratch on 33.5 trillion tokens with a 128K token training context, extendable to 1M tokens for fine-tuning and inference. Comparing MoE model specs before picking one to fine-tune gets easier when tracked on daily.dev."}},{"@type":"Question","name":"How does ERNIE-5.5-Flash-Base compare to DeepSeek-V4-Flash-Base on coding benchmarks?","acceptedAnswer":{"@type":"Answer","text":"ERNIE-5.5-Flash-Base outperforms DeepSeek-V4-Flash-Base on code benchmarks under a 1-shot evaluation setup. It scores 82.3% on HumanEval+ versus 69.5% for DeepSeek-V4-Flash-Base, and 40.1% versus 31.9% on LiveCodeBench pass@1. Developers weighing base models for code generation can follow benchmark comparisons like this on daily.dev."}},{"@type":"Question","name":"Does vLLM support Baidu's ERNIE-5.5-Flash-Base model out of the box?","acceptedAnswer":{"@type":"Answer","text":"Yes, vLLM supported ERNIE-5.5-Flash-Base from day zero through Baidu's vLLM-ERNIE plugin running on vLLM 0.32.0. The model runs in FP8 or BF16, with recipes covering expert parallelism, KV offload, and prefill/decode disaggregation, verified on NVIDIA GB200 and H100 hardware. Teams deploying new MoE checkpoints can track day-0 serving support like this through daily.dev."}}]} ```