ERNIE

Start with Ernie 4.0 via Baidu Qianfan API — Ernie is Baidu's closed-weight flagship family with the strongest Chinese-language performance of any model (C-Eval 93.2%, CMMLU 91.8%). There is no open-weight Ernie variant for local deployment. The Ernie 4.0 Turbo variant offers 5× faster API response at 20% lower cost than Ernie 4.0 with ~3% benchmark degradation — use this for production Chinese NLP pipelines. For open-weight alternatives with strong Chinese performance, use Qwen 3 32B (Apache 2.0) or Yi-1.5 34B (Apache 2.0) — both achieve ~85% of Ernie 4.0's Chinese benchmark scores at zero API cost. Ernie Bot (Baidu's ChatGPT equivalent) provides web search, plugin ecosystem, and iRAG (image RAG) capabilities that open-weight models cannot replicate. The Ernie family's key architectural features (knowledge-enhanced pretraining, heterogeneous multi-expert MoE, and RLHF with constitutional AI) are described in Baidu technical papers but weights remain proprietary.

Ernie is API-only — there is no self-hosted deployment path. Baidu Qianfan API: Ernie 4.0 8K ($0.02/1K input, $0.08/1K output tokens), Ernie 4.0 Turbo 128K ($0.003/1K input, $0.012/1K output), Ernie Speed 8K ($0.001/1K input, $0.004/1K output). For Chinese-market production: Ernie 4.0 Turbo behind a Qianfan-managed endpoint with auto-scaling — latency 200-500ms for typical Chinese queries. For global deployments: Qianfan API has latency penalties outside China (400-800ms to US West, 600-1000ms to Europe) — deploy Ernie in China-region Qianfan for domestic users; use Qwen self-hosted for international users. Ernie's function-calling API is fully compatible with OpenAI function-calling format. For on-premise Chinese NLP where API is not acceptable, self-host Qwen 3 32B with vLLM AWQ 4-bit on H100 SXM — the quality gap vs Ernie 4.0 Turbo is ~8% on C-Eval.