ERNIE 4.5 21B A3B PT — Hardware Requirements & GPU Compatibility
Chat
ERNIE 4.5 21B A3B PT is a 21B-parameter open language model from Baidu in the ERNIE family. It supports a context window of up to 131,072 tokens. At Q4_K_M it needs about 13.02 GB of VRAM — see which GPUs and Macs can run it below.
Runs great on a 24 GB GPU
Runs from 8 GB (IQ2_S · 16K)Best on 24 GB (Q4_K_M · ~254 tok/s · 128K)Macs from 16 GB (Q3_K_L · ~66 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run ERNIE 4.5 21B A3B PT
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
ollama run hf.co/unsloth/ERNIE-4.5-21B-A3B-PT-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 10.2 GB free · 16K context
Download: ERNIE-4.5-21B-A3B-PT-Q4_K_M.gguf13.33 GBunsloth
Other ways
each opens step-by-stepRun in cloud
Fits on RTX 3090 (24 GB) (10 GB headroom) · Q4_K_M
- Generation speed
- ~243 tok/s
- generation speed
- Cost per 1M output tokens
- $0.25
- per 1M output tokens
How Much VRAM Does ERNIE 4.5 21B A3B PT Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 9.3 GB | 11.1 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 128K context · ~307 tok/s | |
| 9.6 GB | 11.4 GB | 3-bit small — 3-bit small quantization | Fits: 128K context · ~303 tok/s | |
| 10.7 GB | 12.4 GB | 3-bit medium — 3-bit medium quantization | Fits: 128K context · ~286 tok/s | |
| 10.9 GB | 12.7 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 128K context · ~282 tok/s | |
| 13.0 GB | 14.8 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 128K context · ~254 tok/s | |
| 15.4 GB | 17.1 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 128K context · ~229 tok/s | |
| 17.7 GB | 19.5 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 64K context · ~208 tok/s | |
| 21.4 GB | 23.2 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 32K context · ~183 tok/s |
Selected Q4_K_M
4-bit medium quantization13.8GBat 16K context
- Weights
- 12.6 GB
- KV cache
- 0.9 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 10.2 GB free · ~254 tok/s
≥50 tok/s needs ≥130 GB/s · RTX 5090 ~347, RTX 3090 Ti ~254
Which Hardware Can Run ERNIE 4.5 21B A3B PT?
Q4_K_M · 13.0 GBERNIE 4.5 21B A3B PT (Q4_K_M) requires 13.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 17+ GB is recommended. Using the full 131K context window can add up to 7.4 GB, bringing total usage to 20.4 GB. 26 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 124 devices with unified memory can run ERNIE 4.5 21B A3B PT, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), MacBook Pro 14" M3 Pro (18 GB).
16–18 GB
16–18 GB tier →fits · 2.2 GB free · up to 32K ctx
18 GPUs · 1 Mac class
RTX 5080~247Simulate output on RTX 5080RTX 5070 Ti~236Simulate output on RTX 5070 TiRTX 4080 SUPER~208Simulate output on RTX 4080 SUPERM3 Pro*18 GB~49Simulate output on MacBook Pro 14" M3 Pro (18 GB)20 GB
fits · 6.2 GB free · up to 64K ctx
1 GPU
RX 7900 XT~113Simulate output on RX 7900 XT32–40 GB
fits · 18.2 GB free · up to 128K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~347Simulate output on RTX 5090M5 Max36 GB~96Simulate output on Mac Studio M5 Max (36 GB)M532 GB~50Simulate output on MacBook Air 13" M5 (32 GB)Orin~75Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~48Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to ERNIE 4.5 21B A3B PT
ERNIE family · 6 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| InternVL3 5 GPT OSS 20B A4B Preview HFOpenGVLab | — | 13.6 GB | ~180 | same size |
| Internlm 20BInternLM | — | 45.3 GB | —needs > 24 GB | same size |
| Tri 21B Thinktrillionlabs | — | 45.1 GB | —needs > 24 GB | same size |
| EuroLLM 22B Instruct 2512utter-project | — | 17.5 GB | ~46 | same size |
| Mistral Small Instruct 2409Mistral AI | — | 17.4 GB | ~46 | same size |
| GPT Neox 20BEleutherAI | — | 13.7 GB | ~48 | same size |
Specifications
- Publisher
- Baidu
- Family
- ERNIE
- Parameters
- 21B
- Active parameters
- 2.5B per token
- Architecture
- Ernie4_5_MoeForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 103,424
- Release Date
- 2025-06-28
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does ERNIE 4.5 21B A3B PT need?
ERNIE 4.5 21B A3B PT requires 13.0 GB of VRAM at Q4_K_M, or 42.4 GB at BF16. Full 131K context adds up to 7.4 GB (20.4 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 21B × 4.8 bits ÷ 8 = 12.6 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 7.8 GB (at full 131K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run ERNIE 4.5 21B A3B PT?
Yes, at Q8_0 (21.4 GB) or lower. Higher quantizations like BF16 (42.4 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for ERNIE 4.5 21B A3B PT?
For ERNIE 4.5 21B A3B PT, Q4_K_M (13.0 GB) offers the best balance of quality and VRAM usage. Q4_K_L (13.3 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 6.2 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run ERNIE 4.5 21B A3B PT on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run ERNIE 4.5 21B A3B PT, the smallest at Q3_K_L. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 13.0 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).
Can I run ERNIE 4.5 21B A3B PT locally?
Yes — ERNIE 4.5 21B A3B PT can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is ERNIE 4.5 21B A3B PT?
At Q4_K_M, ERNIE 4.5 21B A3B PT can reach ~169 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~254 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 1.6 GB active ÷ (8000 × 0.65) = 0.30 ms, plus 28 layers × 0.055 ms = 1.54 ms, so 1000 ÷ 1.84 ms = ~543 tok/s
Estimated speed at Q4_K_M (13.0 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of ERNIE 4.5 21B A3B PT?
At Q4_K_M, the download is about 12.60 GB. The full-precision BF16 version is 42.00 GB. The smallest option (IQ2_XXS) is 5.78 GB.
Which GPUs can run ERNIE 4.5 21B A3B PT?
26 consumer GPUs can run ERNIE 4.5 21B A3B PT at Q4_K_M. Top options include AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, AMD Radeon RX 6800. 8 GPUs have plenty of headroom for comfortable inference.
Which devices can run ERNIE 4.5 21B A3B PT?
124 devices with unified memory can run ERNIE 4.5 21B A3B PT at Q4_K_M (13.0 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.