Baidu Ships ERNIE 5.5 Flash Preview, a Post-Trained 280B MoE on Its Raw Base, API-First With Apache 2.0 Weights Promised
Updated Nov. 12: Baidu on November 11 released ERNIE 5.5 Flash Preview, the first post-trained model built on ERNIE 5.5 Flash Base, with 280 billion total and 14 billion active parameters and a 1 million-token context. The API is available on Baidu's Qianfan platform, and Baidu says the weights will be open-sourced under Apache 2.0 soon; TokenPost reports before November 25. Baidu calls the Preview an intermediate post-training checkpoint. Earlier: on November 5 Baidu released the Base checkpoint itself, pretrained on 33.5 trillion tokens, under its own licence.
Updated November 12, 2026: Baidu has released ERNIE 5.5 Flash Preview, the post-trained model this base was built for. The new section directly below covers it; the November 7 report follows unchanged.
November 12: the post-trained Preview arrives, API first#
The follow-up Baidu promised came within a week. On November 11 it released ERNIE 5.5 Flash Preview, which it calls "the first post-trained model built on the ERNIE 5.5 Flash Base." It keeps the base's shape: 280 billion total parameters, 14 billion active and a 1 million-token context. Baidu says it is strong at coding, 3D modeling, knowledge work and content creation, and scores "competitively across major benchmarks while keeping decoding fast."
Access is API-first. The API is available on Baidu's Qianfan platform, and Baidu says the weights "will be open-sourced under Apache 2.0 soon." TokenPost reports the weights are due before November 25; Baidu's post gives no date. Apache 2.0 would be a looser licence than the base checkpoint's own agreement described below.
Baidu's showcase leans on long agentic jobs, including a third-person Unreal Engine 5.8 game it says the model wrote in C++ in a single 11-hour coding task inside an agent harness, and nearly 100 mobile mini-games built from design documents. Baidu also says the Preview is "an intermediate checkpoint from post-training" and that "some capabilities may still fall short" before an official release. Alibaba took a similar preview-first route this week with Qwen4-Max, offered by subscription only.
Limits of the November 12 update#
- Benchmark and capability claims are Baidu's; we found no independent leaderboard result we could check, and posts on X citing arena scores were not verified.
- The weights are not yet published, so the Apache 2.0 licence is a promise, not a release.
- The demonstrations were chosen by Baidu and used external assets it credits.
Baidu is publishing the raw pretrain, not a chat model, and asking other people to do the expensive part that comes next. On November 5 it released ERNIE 5.5 Flash Base, a sparse mixture-of-experts language model with 280 billion total parameters and 14 billion activated per token. The company says the model was pretrained from scratch on 33.5 trillion tokens, that training context was extended to 128,000 tokens, and that fine-tuning and inference support a 1 million-token context. The weights are the checkpoint at the end of pretraining, before any mid-training or post-training.
That distinction matters for anyone comparing it with instruction-tuned peers. Baidu positions the release as a shared starting point for domain adaptation, instruction tuning and reinforcement-learning research, and says a post-trained ERNIE 5.5 Flash will follow. The weights are on Hugging Face in FP8 and BF16 safetensors under the ERNIE 5.5 Flash Base Model License Agreement v1.0, not Apache or MIT. Baidu says SGLang and vLLM serve the model through the completions API.
How Baidu scores it#
The company evaluated the checkpoint against three other recently released open base models of similar size — DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base — using, it says, the same harness, prompts, few-shot examples and LLM judge. In that setup Baidu reports, among other figures, MMLU at 93.3 versus 89.4 for DeepSeek-V4-Flash-Base, HumanEval+ pass@1 at 82.3 versus 69.5, and LiveCodeBench pass@1 at 40.1 versus 31.9. Those are vendor-run base-model scores, not an independent leaderboard snapshot.
Architecturally, Baidu describes a hybrid of full-context and heavily compressed attention, a multi-router latent MoE, and several other efficiency choices aimed at further training and long context. We are treating those descriptions as the lab's account of its own design.
Limits#
- All benchmark figures are from Baidu's controlled comparison; we have not re-run them or found a matching independent board entry this run.
- The checkpoint is pretrained only. Chat quality, tool use and refusal behavior will depend on whoever post-trains it.
- The license is Baidu's own agreement, not a standard permissive open-source license; downstream users need to read it.
- "Supports a 1 million-token context for fine-tuning and inference" is Baidu's claim; we have not measured long-context quality at that length.