Baidu Research has released the open MoE model ERNIE-5.5-Flash-Base: with 280B total parameters, only 14B are active per token, approximately five percent. The context window for post-training and inference has been extended to 1M tokens, and weights are already available in FP8 and BF16 on Hugging Face with support for SGLang and vLLM. According to vendor benchmarks, the model outperforms DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base in most categories, although only the base version without instruct-tuning has been released so far.
What happened
Baidu Research published the ERNIE-5.5-Flash-Base model in open access along with a technical release description. Training was conducted on 33.5T tokens, pre-training was performed with a 128K window, and the context was expanded to 1M tokens during post-training and inference. The architecture includes hybrid Latent-Gated Sparse Attention, combining MLA with a compressed HCA mechanism, and a Multi-Router Latent MoE block with 512 experts, where ten main and one shared expert are selected per token. A separate ECHO mechanism reuses the model's last layer to predict multiple tokens ahead and simultaneously serves as a built-in draft model for speculative decoding, eliminating the need for a separate compact model. Reported results: MMLU 93.3, GPQA 60.2, HumanEval+ 82.3, LongBench-v2 48.7. Weights are uploaded in FP8 and BF16 formats, with support for SGLang and vLLM announced from day one.
Context
Sparsity is the key idea of the release: the total model size determines its quality and memory footprint, while per-token computation only occurs on the active part, so 14B active parameters out of 280B promise cheap inference. The ratio of training to efficiency is notable: approximately 2,400 pre-training tokens per active parameter, indicating dense training of the active part of the model. The second element is KV-cache management: hybrid attention with compression allows maintaining a million-token context with reasonable memory costs, which is critical for agentic scenarios and processing large documents where models need to hold a lot of text in the working field simultaneously. Finally, the Base suffix means a fundamental model without instruct-tuning: it is not ready for the role of a chat assistant, but is suitable for continued pretraining and fine-tuning for specific tasks, while the full ERNIE 5.5 Flash with product behavior is expected later.
Why this matters for the industry
For companies, this is a cheap entry point into self-hosted long-context products: agentic pipelines and large document processing can be built on own hardware, and the open 280B MoE model can be fine-tuned for specific tasks without purchasing a proprietary API. Architectural techniques—multiple routers for precise expert selection, compressed attention for KV-cache, and speculative decoding without a separate draft model—show concrete ways to reduce inference costs while maintaining quality; if independent evaluations confirm this, these techniques will likely become common in open models, and the combination of hundreds of billions of total parameters with single-digit to tens of billions of active parameters and multi-million context will become a benchmark for the quality-to-cost ratio. For startups, there is another takeaway: access to a strong base model is no longer an advantage, and defense will now have to rely on data, vertical fine-tuning, and distribution, rather than model selection.
Why this matters for users
Technical readers can already download the base model weights in FP8 or BF16, spin up an endpoint on SGLang or vLLM, and run a PoC on a 128K context. The most realistic scenario at this moment is experiments, continued pretraining, and fine-tuning on own data, as well as inference for internal tasks: in a few days, a team can measure the cost-to-quality ratio on their own tasks, including long documents, and compare the result with proprietary APIs. ECHO accelerates generation through built-in speculative decoding, making the model more convenient to test in interactive scenarios. For product chat tasks, the model is still premature: only the Base version without instruct-tuning has been released, and there are no ready-made chat builds or third-party API services.
What is currently unknown / limitations
All reported benchmarks are vendor-provided, the evaluation protocol is not disclosed in the release materials, and the phrasing "outperforms in most categories" itself allows for a wide variance, so the numbers cannot be counted as proof of superiority over DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base, and LongBench-v2 remains a vendor measurement used to verify the claim of strong long context. The claimed speedup from ECHO is plausible by design, but the actual speed increase and the quality of draft-model alignment with the main model are not confirmed in the release. Key operational data is missing: no latency and throughput metrics, nor hardware requirements. Only the Base version has been released, and the timeline for the full ERNIE 5.5 Flash with instruct-tuning is currently unknown, so until its release, it is more practical to limit activities to experiments, fine-tuning, and re-verifying metrics on own tasks.
Sources
Author
Look at AI, editorial team