{"slug": "baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25", "title": "Baidu Launches ERNIE 5.5 Flash Preview, Promises Open Weights by November 25", "summary": "Baidu unveiled ERNIE 5.5 Flash Preview at its Baidu World 2026 conference on November 12, a sparse mixture-of-experts model with 280 billion total parameters and 14 billion activated per token that supports a 1-million-token context window. The preview is available immediately via API on Baidu AI Cloud's Qianfan platform, priced at RMB 1 per million input tokens, RMB 4 per million output tokens and RMB 0.04 per million cached input tokens, and free on OpenRouter, with Baidu saying it will open-source the model weights by November 25. The preview is the first post-trained model built on ERNIE 5.5 Flash Base, which Baidu open-sourced on November 5 after training it from scratch on 33.5 trillion tokens with a 128K training context, though the base weights carry a custom license excluding the EU, UK and South Korea and banning military use.", "body_md": "# Baidu Launches ERNIE 5.5 Flash Preview, Promises Open Weights by November 25\n\nAt its Baidu World 2026 conference on November 12, Baidu unveiled ERNIE 5.5 Flash Preview, a sparse mixture-of-experts model with 280 billion total parameters and 14 billion activated per token. The model supports a 1-million-token context window. Baidu says it performs strongly in coding, 3D modeling, knowledge work, and content creation [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). A third-party aggregator adds that the model is positioned for long-horizon tasks that use planning, tool use and iterative refinement [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview).\n\n## A post-trained sibling to the November 5 base release\n\nERNIE 5.5 Flash Preview builds on ERNIE 5.5 Flash Base, a pre-training checkpoint Baidu open-sourced on November 5. The base model has the same 280B/14B architecture, was trained from scratch on 33.5 trillion tokens with a 128K training context, and supports 1M tokens for fine-tuning and inference [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe)[\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). According to a third-party aggregator, the preview is \"the first post-trained model built on ERNIE 5.5 Flash Base\" and an \"intermediate checkpoint from post-training that is still being improved, and some capabilities may still fall short\" [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview).\n\n## API available now; weights promised this month\n\nThe preview model is accessible immediately via API on Baidu AI Cloud's Qianfan platform, and for free on OpenRouter [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). A third-party aggregator lists the model ID as `ernie-5.5-flash-preview` [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview). Pricing on Qianfan is RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.04 per million cached input tokens [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). Baidu says it will open-source the model weights by November 25 [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). The same aggregator reports the weights are announced for release under Apache 2.0 [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview); Baidu's press release and base model blog do not specify the license for the preview model [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html)[\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base). The base model weights are already on Hugging Face and ModelScope in FP8 and BF16 safetensors, but under a custom ERNIE 5.5 Flash Base Model License Agreement v1.0 that excludes the EU, UK and South Korea and bans military use and using outputs to improve other AI models [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base).\n\n## Performance claims are self-reported\n\nBaidu says ERNIE 5.5 Flash Preview achieves \"competitive scores while maintaining extremely fast decoding speed\" on Terminal-Bench and AutomationBench [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). For the base model, Baidu published a 30-benchmark evaluation against DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B), all run through Baidu's unified pipeline with the same prompts and few-shot settings [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe)[\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). The base model leads on 24 of 30 benchmarks, including MMLU-Pro (78.4 vs 69.2), HumanEval+ (82.3 vs 69.5) and LongBench-v2 (48.7 vs 44.5), but trails on six benchmarks: ChineseSimpleQA, INCLUDE-lite, HELMET-Rerank, InfiniteBench, BBH (CoT) and FullStackBench [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). Baidu notes that competitor numbers in its pipeline differ slightly from vendors' own reports [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe).\n\n## Demos target developers and creators\n\nA third-party aggregator lists official demos including game development in Unreal Engine, Unity and Godot; procedural 3D work in Blender and FreeCAD; research reports and spreadsheets; and tool-assisted video production [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview). The aggregator notes the model is text-only and does not see the screen natively, using tools when visual understanding is required [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview). Baidu says the preview is optimized with ERNIE Evolve, its proprietary recursive self-improvement technology, \"to boost inference efficiency and reduce cost\" [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html).\n\n## Architecture highlights from the base model\n\nThe underlying architecture uses a hybrid attention scheme: 37 of 43 layers employ Heavily Compressed Attention (HCA), caching one entry per 128 tokens with 512-dimension heads; the remaining six layers use Multi-head Latent Attention with sparse attention (MLA-DSA) at 256-dimension heads, keeping token-level access [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe)[\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). Each MoE layer has 512 routed experts (top-10) plus one shared expert, with multiple independent sigmoid routers—a \"Multi-Router\" design Baidu says improves correlation between routing scores and actual expert utility, especially in middle and deep layers [\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). Experts operate in a 2048-dimension latent space, half the hidden width, cutting dispatch/combine traffic and expert FLOPs roughly in half [\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). The model replaces multi-token prediction modules with ECHO, which reuses the backbone's final layer to predict future tokens and doubles as a built-in draft model for speculative decoding [\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base). Pre-training used the Muon optimizer with FP8 mixed precision and a distributed Muon implementation that fuses dense and MoE parameter synchronization across thousands of GPUs [\\[6\\]](https://ernie.baidu.com/release/ernie-5-5-flash-base)[\\[7\\]](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base).\n\n## Licensing split raises questions\n\nThe base model's custom license requires commercial users with over US$50 million in related annual revenue to seek authorization, and excludes the EU, UK and South Korea [\\[1\\]](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe). The preview model's promised Apache 2.0 license would be more permissive, but the weights are not yet public [\\[4\\]](https://theresanaiforthat.com/model/ernie-5-5-flash-preview). The preview announcement does not name a spokesperson or provide a direct quote from leadership about the preview model specifically [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html).\n\n## Part of a broader \"intelligent organization\" push\n\nThe ERNIE 5.5 Flash Preview launch accompanied Baidu's Token Exchange marketplace for model services, upgrades to agents DuMate, KOOKO, Baidu Yijing and Famou 3.0, and hardware announcements including the Tianchi 1024-card Supernode and Kunlunxin M100/M300 chip roadmap [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html). CEO Robin Li framed the moves around three principles: breaking job boundaries, capturing tacit knowledge as a shared asset, and designing closed-loop processes for continuous evolution [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html).\n\n## What remains unverified\n\nIndependent evaluations of the preview model do not yet exist. The base model benchmarks are self-reported and run in Baidu's pipeline. The November 25 open-weight target for the preview model is a timeline, not a delivery. Real-world performance on the demoed tasks (Unreal Engine, Blender, video production) has not been independently tested. The claim that ERNIE Evolve boosts inference efficiency and reduces cost is not quantified in the release [\\[5\\]](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html).\n\n## Sources\n\n1. [Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE ...](https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe) (marktechpost.com)\n2. [ERNIE 5.5 Flash Base: 280B MoE, 14B Active \\[2026\\]](https://shattered.io/baidu-ernie-5-5-flash-base-280b-moe-14b-active-2026) (shattered.io)\n3. [ERNIE 5.5 Flash Base](https://baike.baidu.com/en/item/ERNIE%205.5%20Flash%20Base/7517754) (baike.baidu.com)\n4. [ERNIE 5.5 Flash Preview | AI Model](https://theresanaiforthat.com/model/ernie-5-5-flash-preview) (theresanaiforthat.com)\n5. [Baidu Unveils Latest ERNIE Models Amid Broader Full-Stack AI Lineup, and Debuts Token Exchange to Power Intelligent Organizations](https://www.prnewswire.com/news-releases/baidu-unveils-latest-ernie-models-amid-broader-full-stack-ai-lineup-and-debuts-token-exchange-to-power-intelligent-organizations-302937959.html) (prnewswire.com)\n6. [ERNIE 5.5 Flash Base · ERNIE](https://ernie.baidu.com/release/ernie-5-5-flash-base) (ernie.baidu.com)\n7. [baidu/ERNIE-5.5-Flash-Base · Hugging Face](https://huggingface.co/baidu/ERNIE-5.5-Flash-Base) (huggingface.co)", "url": "https://wpnews.pro/news/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25", "canonical_source": "https://epyhia.com/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by", "published_at": "2026-11-12 17:23:53+00:00", "updated_at": "2026-11-12 17:54:30.801316+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-products", "ai-tools"], "entities": ["Baidu", "ERNIE 5.5 Flash Preview", "ERNIE 5.5 Flash Base", "Baidu AI Cloud", "Qianfan", "OpenRouter", "Hugging Face", "ModelScope"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25", "markdown": "https://wpnews.pro/news/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25.md", "text": "https://wpnews.pro/news/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25.txt", "jsonld": "https://wpnews.pro/news/baidu-launches-ernie-5-5-flash-preview-promises-open-weights-by-november-25.jsonld"}}