Skip to main content

ERNIE · Reasoning · ERNIE 4.5 Thinking

ERNIE 4.5 21B-A3B Thinking: Specs, GPU requirements, hosting & deployment

ERNIE 4.5 21B-A3B Thinking is a current Reasoning release in the ERNIE family, certified from its immutable official source revision.

Technical deployment reference illustration for ERNIE 4.5 21B-A3B Thinking
Reference deployment flow; exact architecture facts are listed separately.
Total parameters21.0B
Active parameters3.0B
Context window131,072 tokens
Release statusCurrent · Current
ArchitectureErnie4_5_MoeForCausalLM
ModalitiesText
Licenseapache-2.0
Deployment stateDeployable with certified configuration
Commercial stateCommercial Use Allowed
Release date2025-09-08

Architecture, strengths and use cases

ERNIE 4.5 21B-A3B Thinking is an official self-hostable Baidu release in the ERNIE Reasoning track. It uses Ernie4_5_MoeForCausalLM, carries 21B / 3B MoE, supports a verified context window of 131,072 tokens, and exposes only source-certified checkpoints and runtime combinations. Its release, license and deployment facts are tied to the official repository revision rather than inferred from the family name.

ERNIE 4.5 21B-A3B Thinking uses Ernie4_5_MoeForCausalLM with 21B / 3B MoE. The stored architecture profile records 28 layers, hidden width 2560, 20 attention heads and a 131,072 token context ceiling where the official configuration publishes those fields.

Verified Ernie4_5_MoeForCausalLM architecture
131,072 token certified context
Official BF16, Unknown checkpoint coverage
Private LLM inference
Agentic workflows
Long-context document processing

Certified checkpoints and quantizations

Only exact, source-certified variants attached to this release are shown.

ERNIE-4.5-21B-A3B-Thinking

baidu/ERNIE-4.5-21B-A3B-Thinking

Official · BF16 · safetensors

Runtime certified: Fastdeploy vLLM Transformers

Precision / quantizationFormatBitsPublisher trustDeployment state
BF16safetensors16OfficialRuntime certified

Recommended hardware

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

Estimated weight memory40.29 GB
KV cache0.44 GB
Runtime overhead12.05 GB
Total with safety margin65.44 GB

Recommended best value

1× NVIDIA H200 SXM 141GB

141 GB total VRAM · nvswitch

75.56 GB headroom · Limited capacity

USD 2,345 / month

Lowest-cost viable

3× NVIDIA L4 24GB

72 GB total VRAM · pcie

6.56 GB headroom · Available

USD 750 / month

Highest performance

1× NVIDIA H200 SXM 141GB

141 GB total VRAM · nvswitch

75.56 GB headroom · Limited capacity

USD 2,345 / month

License and self-hosting considerations

The full 21B 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 21B-A3B Thinking is published under apache-2.0. The release-scoped official license link and verification date are retained with the catalogue record.

Commercial Use Allowed. The official license permits commercial use; its obligations still apply.

Release timeline

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

Source details

Official source

Source revision: aa54548a750802c4c7471b17b3a3b850f3f4b0465a2cf4c43f0d1dd9c9dc5297

Frequently asked questions

How much VRAM does ERNIE 4.5 21B-A3B Thinking require?

The exact estimate depends on checkpoint precision, context and concurrency; the page recommendation uses the certified memory engine and compatible runtime configurations.

Which runtimes support ERNIE 4.5 21B-A3B Thinking?

Certified upstream runtime evidence is available for fastdeploy, vllm, transformers.

Can ERNIE 4.5 21B-A3B Thinking be used commercially?

The stored commercial classification is Commercial Use Allowed. Consult the official apache-2.0 terms for the final obligations.

Official references

Last verified: 2026-09-03