Baidu ERNIE 4.5 Turbo โ flagship reasoning model with 128K context.
ERNIE 4.5 Turbo (128K) is Baidu's efficient, cost-optimised text model in the ERNIE 4.5 family, open-sourced June 30, 2025 under Apache 2.0. It corresponds to the ERNIE-4.5-21B-A3B variant โ a 21B-total-parameter MoE with 3B active parameters per token. Despite its smaller active footprint relative to the flagship 300B-A47B, the Turbo variant inherits the same heterogeneous MoE pre-training pipeline built on PaddlePaddle and the same 128K context window.
For TheRouter operators, ERNIE 4.5 Turbo is the practical entry point into the ERNIE 4.5 generation for latency-sensitive or high-volume Chinese-language workloads. At $0.12 / $0.48 per 1M tokens (input / output), it sits significantly below its sibling baidu/ernie-4.5 in cost while sharing the same OpenAI-compatible API surface. It is well-suited to general Chinese Q&A, instruction-following agents, and structured-output tasks that do not require the full parameter budget of the 300B flagship.
Baidu's own Qianfan platform serves ERNIE 4.5 Turbo via its proprietary REST API and SDK, with pricing in CNY (0.0008 / 0.0032 per 1K tokens input / output as of 2026-07-29). The vendor API uses a different authentication scheme and endpoint structure from OpenAI. Context window is 128K; max output is 8K per the vendor model card.
TheRouter exposes ERNIE 4.5 Turbo as an OpenAI-compatible /v1/chat/completions endpoint under model id baidu/ernie-4.5-turbo-128k. Pricing is converted from CNY at the FX rate recorded in standard-models.yaml ($0.12 input / $0.48 output per 1M tokens as of 2026-07-29). Authentication uses the standard Bearer token scheme. Supported parameters (temperature, max_tokens, top_p, tools, tool_choice, response_format, stop) are taken from the live TheRouter catalog and are a subset of Qianfan's native capability surface.
| Type | Rate |
|---|---|
| Input | $0.1296 / 1M tokens |
| Output | $0.5184 / 1M tokens |
| Release date | 2025-06-30 (open source)ernie.baidu.com โ | verified |
| Architecture | 21B total / 3B active MoE (heterogeneous modality)huggingface.co โ | verified |
| Context window | 128K tokensgithub.com โ | verified |
| License | Apache 2.0 (weights & code)ernie.baidu.com โ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| Inference backends | FastDeploy (PaddlePaddle), SGLang, vLLM; FP8/BF16/INT4/INT2; NVIDIA + AMD + Huawei Ascendgithub.com โ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
C-Eval ERNIE-4.5-21B-A3B-Base reported as competitive against same-class models on Chinese general knowledge; exact per-variant figures not publicly isolated from the 300B results in official release. | โ | Not publicly disclosed | โ |
MMLU Baidu reports the 21B-A3B variant as highly competitive given its total parameter size (~70% of Qwen3-30B). No isolated MMLU figure published for the Turbo variant. | โ | Not publicly disclosed | โ |
CMMLU The ERNIE 4.5 family leads or matches GPT-4.5-class models on Chinese multitask understanding. Per-variant Turbo figures are not separately published; the 21B-A3B base pre-training uses the same data pipeline as the larger 300B-A47B. | โ | Not publicly disclosed | โ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "baidu/ernie-4.5-turbo-128k",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"baidu/ernie-4.5-turbo-128k","messages":[{"role":"user","content":"่ฏท็ฎ่ฆไป็ปไธไธๆททๅไธๅฎถๆจกๅ๏ผMoE๏ผ็ไผๅฟใ"}]}'Baidu released all 10 variants of ERNIE 4.5 under Apache 2.0, including the efficient ERNIE-4.5-21B-A3B (Turbo) model. The release bundles the ERNIEKit training toolkit and FastDeploy inference engine. The 21B-A3B variant targets high-throughput and cost-conscious deployments while retaining the same 128K context as the full family.
ERNIE 4.5 Turbo is the ERNIE-4.5-21B-A3B variant โ 21B total parameters with 3B active per token โ while the flagship (baidu/ernie-4.5) is the 300B-A47B variant. Turbo delivers faster throughput and lower cost at the expense of some accuracy headroom on hard reasoning and knowledge tasks. Both share the same 128K context window, Apache 2.0 license, and OpenAI-compatible API.
Yes. Weights are published on Hugging Face (baidu/ERNIE-4.5-21B-A3B-Paddle). FastDeploy (PaddlePaddle) provides production-grade PD-disaggregation and 4-bit/2-bit quantization. The model also runs via SGLang and vLLM through community recipes.
It handles English well enough for general tasks, but it is optimised for Chinese-language use cases and benchmarks. For English-first workloads โ especially code generation or complex reasoning โ models like DeepSeek-V3 or Qwen3 variants typically offer better measured performance at comparable prices.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | ernie.baidu.com โ | 2026-08-19 | verified |
| Architecture | huggingface.co โ | 2026-08-19 | verified |
| Context window | github.com โ | 2026-08-19 | verified |
| License | ernie.baidu.com โ | 2026-08-19 | verified |
| Training cutoff | โ | โ | unknown |
| Inference backends | github.com โ | 2026-08-19 | verified |
| C-Eval | ernie.baidu.com โ | 2026-08-19 | unknown |
| MMLU | github.com โ | 2026-08-19 | unknown |
| CMMLU | ernie.baidu.com โ | 2026-08-19 | unknown |
| Baidu open-sources ERNIE 4.5 family including the Turbo (21B-A3B) variant | ernie.baidu.com โ | 2026-08-19 | verified |
| What is the difference between ERNIE 4.5 Turbo and the baidu/ernie-4.5 flagship? | ernie.baidu.com โ | 2026-08-19 | to verify |
| Can I self-host ERNIE 4.5 Turbo? | huggingface.co โ | 2026-08-19 | to verify |
| Is ERNIE 4.5 Turbo good for English tasks? | ernie.baidu.com โ | 2026-08-19 | to verify |