ERNIE 4.5 VL 28B-A3B Thinking: ځانګړنې، د GPU اړتیاوې، کوربه‌توب او ځای‌پرځای کول

معماري، پیاوړتیاوې او کارونې

ERNIE 4.5 VL 28B-A3B Thinking is an official self-hostable Baidu release in the ERNIE Vision track. It uses Ernie4_5_VLMoeForConditionalGeneration, carries 28B / 3B MoE, supports a verified context window of 131,072 ټوکنونه, and exposes only source-certified checkpointونه and د اجرا چاپېریال combinations. Its release, license and deployment facts are tied to the official repository revision rather than inferred from the family name.

ERNIE 4.5 VL 28B-A3B Thinking uses Ernie4_5_VLMoeForConditionalGeneration with 28B / 3B MoE. The stored architecture profile records 28 layers, hidden width 2560, 20 attention heads and a 131,072 ټوکن context ceiling where the official configuration publishes those fields.

تایید شوي checkpoint-ونه او quantization-ونه

یوازې د سرچینې له خوا تصدیق شوې کره بڼې چې له دې نسخې سره تړلې دي، ښودل کېږي.

سپارښتل شوی هارډوېیر

Use only the listed transformers, vllm, fastdeploy د اجرا چاپېریال configuration with the exact checkpoint precision. The recommendation engine checks weight memory, KV cache, د اجرا چاپېریال overhead, safety اضافي ظرفیت, GPU architecture and topology before ranking rentable and reference hardware.

د جواز او خپل کوربه‌توب ملاحظات

The full 28B resident weight set must fit across the selected GPUs. Context and concurrency add KV-cache memory; MoE active parameters do not replace resident-weight memory.

ERNIE 4.5 VL 28B-A3B Thinking is published under apache-2.0. The release-scoped official license link and verification date are retained with the catalogue record.

سوداګریزه کارونه اجازه لري. رسمي جواز سوداګریزه کارونه اجازه کوي؛ د جواز نور مکلفیتونه هم پر خپل ځای دي.

بدیلونه او هم‌نسله بڼې

ERNIE 4.5 VL 424B-A47B

د نسخو مهالوېش

ډېرې پوښتل شوې پوښتنې

رسمي مراجع

• huggingface.co

وروستی تایید: 2026-09-03