ER

Ernie-4.5-21B-A3B-Thinking

ErnieGGUF by unslothUpdated 2026-01-24
Download on Hugging Face
Parameters21.8B
Context128K
Layers28
Architectureernie4_5-moe
Smallest quant6.5 GB
Quick answer

Ernie-4.5-21B-A3B-Thinking needs about 14.3 GB of memory at Q4_K_M (a 13.3 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 27 quants, from 6.5 GB (TQ1_0) to 43.7 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 1.9 GB.

21B ERNIE MoE model focused on advanced reasoning

ERNIE-4.5-21B-A3B-Thinking is a 21 billion‑parameter MoE model with 3 billion active parameters per token. It supports 128K context and is optimized for complex reasoning, tool usage, and long‑form text generation.

Good at

  • Advanced logical reasoning
  • Mathematics and science problem solving
  • Efficient tool‑call integration

Limits

  • Requires 80 GB VRAM for full 21B model
  • No vision capabilities
  • Licensed under Apache‑2.0

Best for

  • Complex reasoning workflows
  • Academic research requiring deep analysis
  • Developing AI assistants with tool integration

Who it’s for. Engineers and researchers building applications that need high‑level reasoning and long‑context understanding. The model is suitable for environments with large GPU memory and a need for function‑call support.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
TQ1_06.5 GB 7.5 GB
Tight Get file →
IQ1_S7.0 GB 8.0 GB
Tight Get file →
IQ1_M7.2 GB 8.2 GB
Too big Get file →
IQ2_XXS7.8 GB 8.7 GB
Too big Get file →
IQ2_M8.0 GB 9.0 GB
Too big Get file →
Q2_K8.2 GB 9.1 GB
Too big Get file →
Q2_K_L8.2 GB 9.1 GB
Too big Get file →
Q2_K_XL8.2 GB 9.2 GB
Too big Get file →
IQ3_XXS9.5 GB 10.5 GB
Too big Get file →
Q3_K_S9.6 GB 10.6 GB
Too big Get file →
Q3_K_M10.5 GB 11.5 GB
Too big Get file →
Q3_K_XL10.6 GB 11.5 GB
Too big Get file →
IQ4_XS11.8 GB 12.8 GB
Too big Get file →
IQ4_NL12.5 GB 13.4 GB
Too big Get file →
Q4_012.5 GB 13.5 GB
Too big Get file →
Q4_K_S12.5 GB 13.5 GB
Too big Get file →
Q4_K_M13.3 GB 14.3 GB
Too big Get file →
Q4_K_XL13.4 GB 14.4 GB
Too big Get file →
Q4_113.8 GB 14.7 GB
Too big Get file →
Q5_K_S15.1 GB 16.1 GB
Too big Get file →
Q5_K_M15.6 GB 16.6 GB
Too big Get file →
Q5_K_XL15.7 GB 16.6 GB
Too big Get file →
Q6_K18.0 GB 18.9 GB
Too big Get file →
Q6_K_XL19.3 GB 20.2 GB
Too big Get file →
Q8_023.2 GB 24.2 GB
Too big Get file →
Q8_K_XL26.6 GB 27.6 GB
Too big Get file →
BF1643.7 GB 44.6 GB
Too big Get file →

For 8 GB VRAM, use Q4_0 (13.5 GB memory at 8k context). For 12 GB VRAM, Q4_K_M (14.3 GB). For 16 GB VRAM, Q5_K_S (16.1 GB). Memory estimates are approximate.

What GPU do you need for Ernie-4.5-21B-A3B-Thinking?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: the best quant that fits is IQ1_S (7.0 GB file, about 8.0 GB in use).
  • 12 GB VRAM: the best quant that fits is Q3_K_XL (10.6 GB file, about 11.5 GB in use).
  • 16 GB VRAM: the best quant that fits is Q4_1 (13.8 GB file, about 14.7 GB in use).
  • 24 GB VRAM: the best quant that fits is Q6_K_XL (19.3 GB file, about 20.2 GB in use).
  • 32 GB VRAM: the best quant that fits is Q8_K_XL (26.6 GB file, about 27.6 GB in use).
  • 48 GB VRAM: the best quant that fits is BF16 (43.7 GB file, about 44.6 GB in use).

Which Ernie-4.5-21B-A3B-Thinking quant should you choose?

Q4_K_M is the usual starting point for Ernie-4.5-21B-A3B-Thinking: it balances file size and quality well. Here it is a 13.3 GB file.

The Q8_0 file is 23.2 GB, about 1.7 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is TQ1_0 at 6.5 GB, with the quality trade-offs that come with fewer bits. Lower bit counts mean smaller files and more quality loss.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architectureernie4_5-moe
Parameters21.8B (approximate, mixture of experts)
Active parameters3B
Layers28
Trained context131,072 tokens
KV cache (fp16)56 KB per token

Source

Quantized byunsloth
FamilyErnie
Licenseapache-2.0
Downloads (HF)7,119
Last updated2026-01-24

Frequently asked questions about Ernie-4.5-21B-A3B-Thinking

How much VRAM does Ernie-4.5-21B-A3B-Thinking need?

About 14.3 GB at Q4_K_M with an 8K context window. That is the 13.3 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Ernie-4.5-21B-A3B-Thinking on an 8 GB GPU?

Yes, with IQ1_S: the 7.0 GB file needs about 8.0 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of Ernie-4.5-21B-A3B-Thinking GGUF?

It depends on the quant: TQ1_0 is 6.5 GB, IQ1_S is 7.0 GB, IQ1_M is 7.2 GB, IQ2_XXS is 7.8 GB, IQ2_M is 8.0 GB, Q2_K is 8.2 GB and 21 more.

What context length does Ernie-4.5-21B-A3B-Thinking support?

The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 0.5 GB at 8K and 1.9 GB at 32K.

Where can I download Ernie-4.5-21B-A3B-Thinking GGUF?

The files are on Hugging Face in the unsloth/ERNIE-4.5-21B-A3B-Thinking-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Ernie-4.5-21B-A3B-Thinking use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run this model on 8 GB VRAM?

Use the Q4_0 quantized file; it requires about 13.5 GB of VRAM at 8k context, so 8 GB is insufficient.

Does the model support vision tasks?

No, the model is text‑only and does not include vision capabilities.

What license applies to this model?

The model is released under the Apache License 2.0, allowing commercial use.

Other GGUFs of ERNIE-4.5-21B-A3B-Thinking

Similar models

Browse all →

More from unsloth

All by unsloth →

More Ernie models

All Ernie →

Other medium models

All medium models →