--- title: Baidu released ERNIE 5.5 Flash Base | daily.dev description: Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a 280B-parameter sparse MoE model with 14B parameters activated per token. It was... canonical: https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Baidu released ERNIE 5.5 Flash Base | daily.dev og:description: Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a 280B-parameter sparse MoE model with 14B parameters activated per token. It was... og:url: https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1 og:image: https://api.daily.dev/og/posts/06EEwzZP1.png og:image:alt: Baidu released ERNIE 5.5 Flash Base 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 released ERNIE 5.5 Flash Base **[Hacker News](https://daily.dev/sources/hn)** · 15 min read · 0 upvotes · 0 comments ## Summary Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a 280B-parameter sparse MoE model with 14B parameters activated per token. It was pretrained from scratch on 33.5T tokens with a training context up to 128K, extendable to 1M tokens for fine-tuning and inference. The architecture mixes Multi-head Latent Attention with Heavily Compressed Attention, uses 512 routed experts plus a shared expert, and adds mHC residual streams, a MultiMax output head and an ECHO module replacing multi-token prediction. It performs strongly on most exam, reasoning and code benchmarks against similarly sized DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base. Weights are on Hugging Face in FP8 and BF16. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How many parameters does ERNIE-5.5-Flash-Base have and how many are activated per token? ERNIE-5.5-Flash-Base is a sparse mixture-of-experts model with 280 billion total parameters, of which 14 billion are activated per token. It was pretrained from scratch on 33.5 trillion tokens with a training context up to 128K tokens, and supports a 1 million token context for fine-tuning and inference. _Developers evaluating base models for fine-tuning can track new MoE releases like this on daily.dev._ ### What is the difference between MLA and HCA attention layers in ERNIE 5.5 Flash? Only one layer in seven uses Multi-head Latent Attention (MLA) with full token-level access to the context, while the rest use Heavily Compressed Attention (HCA), which compresses every 128 tokens into one cached entry plus a short window of recent tokens. This keeps long-context KV cache small while six MLA-DSA layers preserve strong long-context performance. _Engineers comparing efficient long-context attention designs can follow architecture breakdowns like this on daily.dev._ ### How does ERNIE-5.5-Flash-Base compare to DeepSeek-V4-Flash-Base on coding benchmarks? ERNIE-5.5-Flash-Base scores 82.3 Pass@1 on HumanEval+ and 40.1 Pass@1 on LiveCodeBench, versus 69.5 and 31.9 respectively for DeepSeek-V4-Flash-Base (284B-A13B), using the same evaluation harness and few-shot settings. ERNIE also leads on MBPP+ (81.5 vs 67.2) and RepoBench (50.3 vs 49.1). _Teams picking a base model for code-focused fine-tuning can compare benchmark results like these on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 75 comments across hackernews (as of 2026-11-05). **TL;DR:** Much of the discussion is lighthearted wordplay around the ERNIE/BERT/Sesame Street naming, but there's substantive debate about the value of releasing base (non-instruction-tuned) models, the restrictive license excluding the EU/UK/South Korea, and how it stacks up against DeepSeek's Flash base models. **Sentiment:** 30% positive · 45% mixed · 25% skeptical **The case for** - Releasing a raw base model is seen as valuable for the field and for groups wanting to do their own post-training. - Its parameter count and benchmark results are considered a solid, efficient improvement over similarly sized DeepSeek base models. - Some see wide distribution of open base models as a reasonable, decentralized approach to AI alignment risk. **The pushback** - The license explicitly forbids use in the EU, UK, and South Korea, which many see as Baidu offloading legal/regulatory risk onto users. - As a base model it isn't instruction-tuned, so it's not directly usable for coding agents or typical assistant tasks. - Comparisons to DeepSeek are seen as somewhat unfair since Baidu benchmarked against an older DeepSeek-V4-Flash-Base rather than the latest version. - One commenter dismissed it outright as low-quality 'Chinese model slop'. **By community** - hackernews (mixed): A large portion of the thread is joking tangents about the ERNIE/BERT naming, while the more substantive subthreads debate the license restrictions, the merits of releasing base-only models, and how it compares to DeepSeek's flash base models. **Hottest debate:** Whether the EU/UK/South Korea usage restriction in the license is a meaningful legal shield or just an unenforceable, cynical liability dodge. **Open questions** - Does releasing only a base model (without instruction tuning) make this release less immediately useful than fully finetuned competitors? - Is a US court likely to ever rule that model weights are copyrightable, which would affect enforceability of such license restrictions? - What specific EU/UK/South Korea regulations are driving providers to exclude those regions from open-weight releases? **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 · 6 comments](https://news.ycombinator.com/item?id=50378719) > A large base model is ~90% of the effort/cost to making a frontier model, and by itself is not useful for much. There aren't many controversial decisions to make when training a base model - you mostly just train a model as big as possible on as much human-generated data as possible. But it makes it far easier for other parties to do that last 10% to train it to be useful. However, there are many choices one makes in the last 10% of training which will impact the models behaviour when it comes to AI safety, political opinions, willingness to help make a bomb, etc etc. > — [londons\_explore on hackernews · 1 comments](https://news.ycombinator.com/item?id=50378648) > 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) > TLDR: Ernie 5.5 Flash base 280B params 14B active A chunk better than Deepseek 1017, at half the parameter count, BUT it's only a base model, not finetuned to follow instructions. Strange to release a model in this state. OpenAI is also "probably" releasing new models (sources within OpenAI talking about Sol 6.2 and the like). We must be getting really close to the Qwen4 releases. > — [spwa4 on hackernews · 3 comments](https://news.ycombinator.com/item?id=50377335) > 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) · 123 points · 75 comments ## Similar posts on daily.dev - [Intelligence, Performance & Price Analysis](https://daily.dev/posts/intelligence-performance-price-analysis-uowbhxnia) · Hacker News · 10 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1) ```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 released ERNIE 5.5 Flash Base","url":"https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1"},"datePublished":"2026-11-05T12:32:40.657Z","dateModified":"2026-11-05T15:49:04.798Z","description":"Baidu released ERNIE-5.5-Flash-Base, an open base-weights checkpoint for a 280B-parameter sparse MoE model with 14B parameters activated per token. It was...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b9ec8cb67383b329899c3df612ee555d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b9ec8cb67383b329899c3df612ee555d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","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":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,mixture-of-experts","timeRequired":"PT15M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Baidu released ERNIE 5.5 Flash Base"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/baidu-released-ernie-5-5-flash-base-06eewzzp1#faq","mainEntity":[{"@type":"Question","name":"How many parameters does ERNIE-5.5-Flash-Base have and how many are activated per token?","acceptedAnswer":{"@type":"Answer","text":"ERNIE-5.5-Flash-Base is a sparse mixture-of-experts model with 280 billion total parameters, of which 14 billion are activated per token. It was pretrained from scratch on 33.5 trillion tokens with a training context up to 128K tokens, and supports a 1 million token context for fine-tuning and inference. Developers evaluating base models for fine-tuning can track new MoE releases like this on daily.dev."}},{"@type":"Question","name":"What is the difference between MLA and HCA attention layers in ERNIE 5.5 Flash?","acceptedAnswer":{"@type":"Answer","text":"Only one layer in seven uses Multi-head Latent Attention (MLA) with full token-level access to the context, while the rest use Heavily Compressed Attention (HCA), which compresses every 128 tokens into one cached entry plus a short window of recent tokens. This keeps long-context KV cache small while six MLA-DSA layers preserve strong long-context performance. Engineers comparing efficient long-context attention designs can follow architecture breakdowns like this 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 scores 82.3 Pass@1 on HumanEval+ and 40.1 Pass@1 on LiveCodeBench, versus 69.5 and 31.9 respectively for DeepSeek-V4-Flash-Base (284B-A13B), using the same evaluation harness and few-shot settings. ERNIE also leads on MBPP+ (81.5 vs 67.2) and RepoBench (50.3 vs 49.1). Teams picking a base model for code-focused fine-tuning can compare benchmark results like these on daily.dev."}}]} ```