ERNIE-4.5-21B-A3B
Baidu's open-weight MoE flagship from the June 2025 ERNIE 4.5 release: 21B total parameters, only 3B active, Apache 2.0.
About
ERNIE-4.5-21B-A3B is the lightweight mixture-of-experts model Baidu open-sourced on June 30, 2025, as part of a ten-model ERNIE 4.5 release that marked Baidu's biggest open-weights push to date. It totals 21 billion parameters but activates only 3 billion per token, and Baidu reported that the base model beats Qwen3-30B-A3B-Base on multiple math and reasoning benchmarks despite having roughly 30% fewer parameters. The weights shipped under Apache 2.0, with training done on Baidu's PaddlePaddle framework using techniques like router orthogonalization loss and FP8 mixed-precision training.
Baidu followed it with a thinking variant (ERNIE-4.5-21B-A3B-Thinking, September 2025) tuned for deep reasoning with tool use and a 128K context window, and later with ERNIE 5.0. The base model remains a solid pick for teams that want a genuinely permissive Apache 2.0 MoE in the 20B class without the custom-license strings attached to most Chinese open weights. Treat Baidu's benchmark comparisons as vendor-reported — no independent replication of the base model's scores was found.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Baidu-reported: ERNIE-4.5-21B-A3B-Base outperforms Qwen3-30B-A3B-Base on multiple mathematics and reasoning benchmarks (provider claim, ERNIE 4.5 technical report)
- No independent benchmark table verified for the base weights; treat vendor comparisons as provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Baidu's official ERNIE blog announced the open-source ERNIE 4.5 family on Jun 30, 2025; MoE 21B total / 3B active under Apache 2.0; release covered by VentureBeat, MarkTechPost, and Baidu's own PR. Open weights. Strengths: genuinely permissive Apache 2.0 license (rare among Chinese frontier labs), efficient MoE design, active follow-through (Thinking variant Sep 2025, ERNIE 5.0 later). Docked for: (1) headline benchmark claims are Baidu-run with no independent replication found; (2) training data undisclosed. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Baidu's official ERNIE blog announced the open-source ERNIE 4.5 family on 2025-06-30 (10 models, incl. 21B-A3B)
- · MoE: 21B total parameters, 3B active per token; 128K context per Baidu's ERNIE 4.5 series spec
- · Baidu-reported: 21B-A3B-Base outperforms Qwen3-30B-A3B-Base on math/reasoning benchmarks (provider claim)
- · Follow-through: Thinking variant Sep 2025, ERNIE 5.0 later