How much memory does ERNIE-4.5-21B-A3B-Thinking need?

About 14 GB at 4-bit, 25 GB at 8-bit and 46 GB at full 16-bit precision, including an 8,192-token KV cache and about 1.5 GB of runtime overhead. These are estimates from its parameter count.

Can ERNIE-4.5-21B-A3B-Thinking run on a 24 GB graphics card?

Yes, at 4-bit by this estimate, with room left for context.

What is ERNIE-4.5-21B-A3B-Thinking’s context window?

Its configuration sets 131,072 tokens. Longer contexts need more memory for the KV cache, so the practical limit on a given machine can be lower.

ANVE: the private AI office, on the record. See what that means

ANVE

Run it on your own hardware · updated 2026-10-16

ERNIE-4.5-21B-A3B-Thinking
The memory it needs, what you may do with it, and the hardware it suits.

Short answer. ERNIE-4.5-21B-A3B-Thinking needs about 14 GB of memory at 4-bit, 25 GB at 8-bit and 46 GB at full 16-bit precision, including runtime overhead and an 8,192-token KV cache. It has about 21.8 billion parameters and a 131,072-token context window. Licence in short: commercial use allowed. By memory alone, on ANVE’s launch configurations: ANVE One fits at BF16, ANVE Office fits at 4-bit, ANVE Office Pro (GPU) fits at 4-bit. These are estimates from its published size, not benchmarks.

What it is

ERNIE-4.5-21B-A3B-Thinking is a language model for chat, drafting and question answering, published by Baidu on Hugging Face on 8 September 2025. It has about 21.8 billion parameters and a configured context window of 131,072 tokens. It is a mixture-of-experts model with about 3B parameters active per token.

Its card lists the licence as Apache 2.0. In short, the licence allows commercial use, modification and redistribution, with attribution and a copy of the licence, and includes a patent grant.

Hugging Face counted about 85,602 downloads of it in the 30 days before we checked (7 Oct 2026), and 792 people have liked it there.

The card lists 2 languages: en, zh (ISO codes as given on the card).

How much memory it needs

The arithmetic: parameters × bits per weight ÷ 8 gives the weights; add the KV cache, which grows with every token of context; then add about 1.5 GB for the runtime. These are estimates for planning, not measurements.

Each token of context adds about 0.06 MB of KV cache at 16-bit, so 32,768 tokens would add about 1.9 GB.

Precision	Weights	KV cache (8,192 tokens)	Total, estimated
BF16 (full precision)	44 GB	0.47 GB	46 GB
8-bit (Q8_0)	23 GB	0.47 GB	25 GB
4-bit (Q4_K)	12 GB	0.47 GB	14 GB

Architecture at a glance

Setting	Value in config.json
Architecture class	Ernie4_5_MoeForCausalLM
Model type	ernie4_5_moe
Layers	28
Hidden size	2,560
Attention heads	20
Key-value heads	4 (grouped-query attention)
Head size	128
Vocabulary	103,424
Configured context	131,072
Stored precision	bfloat16

Will it fit on ANVE?

Arithmetic against the launch specification, not a test: about three quarters of ANVE One’s 64 GB of unified memory for the model (as AMD’s Variable Graphics Memory allows), and the 24 GB on the graphics card in ANVE Office and ANVE Office Pro, which keep the rest of the stack in 64 GB and 128 GB of system memory. ANVE’s tested default is Qwen 3.8 27B; this model is not yet validated on ANVE.

Configuration (at launch)	Memory for the model	Verdict
ANVE One	about 48 GB of 64 GB unified memory	fits at BF16
ANVE Office	24 GB on the RTX PRO 4000 Blackwell SFF Edition	fits at 4-bit
ANVE Office Pro (GPU)	24 GB on the RTX PRO 4000 Blackwell	fits at 4-bit

What an office would use it for

This is the class ANVE runs by default (27B-class): strong enough for meeting minutes, summaries, drafting and question answering over an office’s own documents, and small enough for one box.

Similar models to compare

EuroLLM-22B-Instruct-2512 (22.6B parameters, about 14 GB at 4-bit before context; commercial use allowed)
gpt-oss-20b (20.9B parameters, about 13 GB at 4-bit before context; commercial use allowed)
internlm2-chat-20b (19.9B parameters, about 13 GB at 4-bit before context; custom licence: check before commercial use)

Questions people ask

How much memory does ERNIE-4.5-21B-A3B-Thinking need?: About 14 GB at 4-bit, 25 GB at 8-bit and 46 GB at full 16-bit precision, including an 8,192-token KV cache and about 1.5 GB of runtime overhead. These are estimates from its parameter count.
Can ERNIE-4.5-21B-A3B-Thinking run on a 24 GB graphics card?: Yes, at 4-bit by this estimate, with room left for context.
Is ERNIE-4.5-21B-A3B-Thinking free for commercial use?: Its card lists Apache 2.0, which allows commercial use, modification and redistribution, with attribution and a copy of the licence, and includes a patent grant. Read the licence itself before relying on this summary; it is not legal advice.
What is ERNIE-4.5-21B-A3B-Thinking’s context window?: Its configuration sets 131,072 tokens. Longer contexts need more memory for the KV cache, so the practical limit on a given machine can be lower.

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

ERNIE-4.5-21B-A3B-ThinkingThe memory it needs, what you may do with it, and the hardware it suits.