ERNIE 4.5 21B A3B

by BaiduChat

Open ERNIE text MoE with 21B total and 3B active parameters

Context
120K tokens
Max output
8,000 tokens
Parameters
21B total, 3B active
Released
2025-06-30
Weights
Open

baidu/ernie-4-5-21b-a3b

Specifications

Full facts as recorded, incl. what is unknown.

Released
2025-06-30
Weights
Open
Input
Text
Output
Text
Context window
120,000 max across endpoints
Max output
8,000
Providers
1
Epoch Capabilities Index
not evaluated
Direct price range
$0.07 in · $0.28 out per 1M tokens · from its one provider
More details
Last updated
2025-06-30
Knowledge cutoff
Parameters
21B total, 3B active
Developer country
China
Training compute
1.8×10²³ FLOP

Capabilities

Across all endpoints: supported if any endpoint supports it.

  • Tool calling
  • Parallel tools
  • Structured outputunknown
  • Reasoning
  • Image input
  • PDF input
  • Audio input
  • Audio outputunknown
  • Video input
  • Prompt cachingunknown
  • Web searchunknown
  • Computer useunknown
  • System messages
  • Assistant prefillunknown
  • Temperature
  • Streamingunknown
  • Attachments

Endpoints

One row per provider and variant; gateways are listed separately.

Loading endpoints…

Price history

No change since 19 Oct 2026. The cheapest direct price has been $0.07 in / $0.28 out per 1M tokens at Novita AI on every data refresh since then.

Tracked from 19 Oct 2026. Headline-eligible direct prices only, so a price the sanity check flags never draws a spike. Series as JSON

Cost calculator

Loading endpoints…

More from Baidu

Newest Baidu models in the catalogue. All Baidu models

ModelReleasedContextListed price
ERNIE 5.1 Thinking2026-05-10119K$0.563 via AIHubMix
CoBuddy2026-05-06131Kfrom $0.28
PaddleOCR-VL-1.52026-01-2916.4Kunpriced
ERNIE 4.5 300B A47B2026-01123Kfrom $0.28
ERNIE 5.02025-11-18128K$0.84 via ZenMux
PaddleOCR VL2025-10-1616.4Kfrom $0.02
ERNIE 4.5 VL 424B A47B2025-06-30123Kfrom $0.42
ernie-4-5-vl-28b-a3b2025-06-2930Kfrom $0.14
Heads up: prices are indicative list rates scraped from provider pricing pages; verify with the provider before relying on them. Report incorrect data for this model