Size
280B
14B active — only part of the model works on each answer, but all of it has to fit in memory
ERNIE 5.5 Flash Base
Context
1M
the longest conversation the model is built to hold
Lab
Baidu
the lab that released it
Released
Nov 5, 2026
Licence: ERNIE 5.5 Flash Base Model License Agreement v1.0 — Baidu’s custom licence: use is authorised only outside the EU, UK and South Korea; non-commercial use is free; commercial use at or below $50M annual revenue from related products/services needs a one-time notice to Baidu plus a “Powered by ERNIE” display, and above $50M needs separate written authorisation; outputs may not be used to improve an unrelated AI model. Read it before commercial self-hosting.
Worth knowing
Baidu’s open-weight base model, released Nov 5, 2026: a sparse MoE checkpoint at the end of pre-training (no mid- or post-training) — 280B total parameters, 14B active per token, 43 layers, 512 routed experts (top-10) plus 1 shared, pre-trained on 33.5T tokens; 128K training context, 1M for fine-tuning and inference. Checkpoints on Hugging Face: FP8 (128 shards, about 304GB) and BF16 (about 573GB). Baidu-reported benchmarks under a unified pipeline (MMLU 93.3, HumanEval+ 82.3) are the maker’s figures, not independently reproduced. At FP8 the weights alone are about 280GB — the card’s serving recipes use four-way parallelism on large cards; BF16 doubles the memory before any training state.
What it needs
Full precision (FP16)
560 GB
the weights, uncompressed
8-bit
280 GB
the weights, compressed to 8-bit
4-bit
140 GB
the weights, compressed to 4-bit
How we work this out: memory = the model's total parameters × bytes per weight — 2 bytes at full precision (FP16), 1 byte at 8-bit, half a byte at 4-bit. Weights only — add room for the conversation memory (KV cache) and the serving software. A GPU counts as running a build when its memory covers the weights plus about 20% headroom. Real figures move with context length and software.
GPUs that can run it
Ranked by how well the model runs, then by price.
Runs ERNIE 5.5 Flash Base compressed — the 4-bit build needs 140GB, which fits this card's 192GB. The compromise: slightly lower answer quality and a shorter conversation memory than the full-precision build.
Frontier training money. If your job fits on anything cheaper, a B200 is budget burned for a spec sheet.
Runs ERNIE 5.5 Flash Base compressed — the 4-bit build needs 140GB, which fits this card's 288GB. The compromise: slightly lower answer quality and a shorter conversation memory than the full-precision build.
Very costly for small jobs — best value only on big work.
At full precision, ERNIE 5.5 Flash Base needs more than one GPU — no single card holds its 560GB of weights plus room to work. Every card above runs a compressed build.
The cheapest verified way to run it
B200 at RunPod Community — $5.98/hr — runs it compressed — the 4-bit build.
Launch RunPod Community ↗ See the B200 →
Checked Nov 15, 2026
Select a GPU See these GPUs on the prices board, ranked for this model.