vLLM adds Day-0 support for Baidu's open ERNIE 5.5 Flash Base, a 280B MoE
vllm_project · x · 2026-11-05
vLLM announces Day-0 support for Baidu's ERNIE-5.5-Flash-Base, the pre-trained base of ERNIE 5.5 Flash before mid/post-training (post-trained version coming soon).
- Architecture: 280B total / 14B active MoE with Multi-Router Latent MoE; 37 compressed-attention layers + 6 MLA-DSA layers; ECHO reuses the final layer as a speculative draft
- Scale: pre-trained on 33.5T tokens; 128K training context, up to 1M for fine-tuning and inference
- Benchmarks: 82.3% HumanEval+ and 40.1% LiveCodeBench pass@1 (1-shot) vs 69.5%/31.9% for DeepSeek-V4-Flash-Base under the same setup
- Serving: FP8/BF16 via Baidu's vLLM-ERNIE plugin on vLLM 0.32.0; recipes cover expert parallelism, KV offload, and prefill/decode disaggregation, verified on NVIDIA GB200 and H100
Open base weights are on Hugging Face for domain adaptation, instruction tuning and RL research.
More from Infra
- 40 of 41 Data Center Ballot Measures in 2026 Midterms Went Against AI Builders — Acceptable-Offer-518 · 2026-11-12
- Azure Day at KubeCon: AKS Runs 22,655 Agents on 65 Nodes in Live Demo — davemccollough · 2026-11-12
- Gas-for-AI-datacenters forum co-located with CoRL shows energy demand spillover — carlosdponx · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- AMD to pass Intel in server revenue by mid-2027 as share gains 6 points a year — BenBajarin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12