{"slug": "baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference", "title": "Baidu open-sources ERNIE 5.5 Flash Base in a race to cheap inference", "summary": "Baidu published the pretraining checkpoint of ERNIE-5.5-Flash-Base on Hugging Face on November 5, a sparse mixture-of-experts model with 280 billion total parameters of which only 14 billion activate per token, released in FP8 and BF16 versions. Baidu trained the 43-layer model from scratch on 33.5 trillion tokens with a 128K training context that it claims stretches to 1 million tokens at inference, and reported base-checkpoint scores of 93.3 on MMLU, 78.4 on MMLU-Pro, 60.2 on GPQA, 82.3 on HumanEval+ and 40.1 on LiveCodeBench. The release landed one day after Moonshot AI opened the weights for Kimi K3.1 Pro on November 4, with Meituan's API-only LongCat-2.5 having arrived on November 2, as the three Chinese labs compete on inference cost per dollar rather than frontier benchmark records.", "body_md": "*Baidu's new ERNIE-5.5-Flash-Base is a sparse model that spends very little compute per token. The company is betting that cheap, fast inference matters more right now than another leaderboard record.*\n\nOf the 280 billion parameters in Baidu's ERNIE-5.5-Flash-Base, only 14 billion activate for any given token. The model is a sparse mixture-of-experts system, and Baidu published its pretraining checkpoint on Hugging Face on November 5. There are two versions, an FP8 repository and a BF16 one, so developers can choose between lower memory use and higher numerical precision. That's the whole point of the Flash label: it is built to run fast and cheap, not to win every benchmark outright.\n\nArchitecturally, ERNIE-5.5-Flash-Base is unusual even by current MoE standards. It stacks 43 layers, one dense feed-forward layer plus 42 Multi-Router Latent MoE layers, and splits attention between 37 Heavily Compressed Attention layers and 6 Multi-head Latent Attention layers that combine MLA with sparse attention. Each MoE layer routes among 512 experts, activating the top 10 per token plus one shared expert that always fires. Baidu trained the whole thing from scratch on 33.5 trillion tokens, with a 128K training context that stretches to a claimed 1 million tokens at inference.\n\nThere's a catch worth sitting with before anyone gets excited about the benchmark numbers: this is a base checkpoint. No post-training, no instruction tuning, no reinforcement learning from human feedback. What Baidu released is the raw end-of-pretraining state, the kind of artifact researchers and model-builders use as a starting point, not the kind consumers chat with directly. Reported scores, MMLU at 93.3, MMLU-Pro at 78.4, GPQA at 60.2, HumanEval+ at 82.3, LiveCodeBench at 40.1, are measured against other base checkpoints, not against finished chat models like GPT-5 or Claude. That context matters because base-model numbers and post-trained numbers aren't the same currency, and conflating them is exactly how benchmark claims get overstated.\n\nThe timing isn't an accident. Baidu's release lands one day after Moonshot AI made the weights for Kimi K3.1 Pro publicly available on November 4, following an API launch on October 23. Moonshot's model is a 3-trillion-parameter-class system with a 1-million-token context window of its own, priced at $2 per million input tokens and $10 per million output on the API, undercutting Western frontier labs the way Moonshot has made a habit of doing. Meituan is in the same race with LongCat-2.5, a 1.6-trillion-parameter multimodal model it introduced on November 2. It has kept that one API-only for now rather than opening the weights the way its earlier LongCat 2.0 did.\n\n[Baidu Says Its AI Business Is About to Earn as Much as Search Does](https://startupfortune.com/baidu-says-its-ai-business-is-about-to-earn-as-much-as-search-does/)\n\nBaidu CFO Haijian He told investors the company's AI business, spanning cloud, chips and apps, is closing in on the profit margins of its legacy search engine, with GPU cloud revenue up 283% in Q2 2026. The claim lands as rivals Z.AI and OpenAI both fall short of their own revenue targets, sharpening the debate over whether AI infrastructure... - [when will AI business become profitable for companies](https://startupfortune.com/baidu-says-its-ai-business-is-about-to-earn-as-much-as-search-does/) - [Baidu AI profit margins matching search engine business](https://startupfortune.com/baidu-says-its-ai-business-is-about-to-earn-as-much-as-search-does/)\n\nThree Chinese labs, three large models, three different bets on how to compete with American frontier labs within weeks of each other. None of them are chasing GPT-5 or Gemini head to head on every benchmark. They're competing on capability per dollar. And on how cheaply a developer can run inference at scale for agentic workloads that fire off thousands of model calls instead of one long chat session. Export restrictions on Nvidia's advanced chips have pushed Chinese labs toward architectures that squeeze more out of less compute, and a 14-billion-active-parameter MoE design is a direct answer to that constraint, not a coincidence.\n\nBaidu's licensing tells its own story about who this release is actually for. The terms exclude the EU, UK, and South Korea outright, and once a deployer's related commercial revenue crosses $50 million, Baidu requires written authorization before continuing to use the model commercially. That's not an open-source licence in the Apache 2.0 sense Baidu used for its ERNIE 4.5 family back in mid-2025. It's a permissive-until-you-succeed arrangement, free enough to seed adoption, restrictive enough that Baidu keeps a hand on the lever once anyone using it starts making real money.\n\nFrankly, that's the more honest signal here than the benchmark table. A lab doesn't write a revenue trigger into its licence unless it expects the model to get used commercially at scale. Baidu is betting developers will build on ERNIE-5.5-Flash-Base specifically because it's cheap to run, and it's hedged that bet by making sure it gets paid if that bet pays off big. Whether the finished, post-trained version lives up to the base checkpoint's numbers is still an open question. Baidu hasn't said when that version ships.\n\n**Also read:** [Amazon is in talks to buy AI startup Decart for about $7 billion](https://startupfortune.com/amazon-is-in-talks-to-buy-ai-startup-decart-for-about-7-billion/) • [Apple cuts AI token budgets for employees as Ternus tightens costs](https://startupfortune.com/apple-cuts-ai-token-budgets-for-employees-as-ternus-tightens-costs/) • [Moonshot AI Releases Full Weights For 2.8 Trillion Parameter Kimi K3.1 Pro](https://startupfortune.com/moonshot-ai-releases-full-weights-for-28-trillion-parameter-kimi-k31-pro/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference", "canonical_source": "https://startupfortune.com/baidu-open-sources-ernie-55-flash-base-in-a-race-to-cheap-inference/", "published_at": "2026-11-05 10:13:24+00:00", "updated_at": "2026-11-05 10:23:09.187832+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-products", "ai-infrastructure", "artificial-intelligence"], "entities": ["Baidu", "ERNIE-5.5-Flash-Base", "Hugging Face", "Moonshot AI", "Kimi K3.1 Pro", "Meituan", "LongCat-2.5", "Haijian He"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference", "markdown": "https://wpnews.pro/news/baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference.md", "text": "https://wpnews.pro/news/baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference.txt", "jsonld": "https://wpnews.pro/news/baidu-open-sources-ernie-5-5-flash-base-in-a-race-to-cheap-inference.jsonld"}}