ERNIE 4.5 21B-A3B
21B-A3BBaidu ERNIE 4.5
Baidu's open mixture-of-experts model: about 21.8B parameters in total and roughly 3B active per token. Each token goes through 6 of 64 routed experts plus 2 shared ones. The total decides the memory and the active count decides the speed, so it needs the VRAM of a 20B-class model and reads about as much per token as a 3B one. At Q4_K_M it needs about 14.2 GB at 4K context, just past the comfortable line on a 16 GB card (it loads, with little margin). A 24 GB card, or a 24 GB Mac through the GPU's share of unified memory, leaves real room for context. Sizes come from the published GGUF files of the PT (instruct) checkpoint; bartowski's Q4_K_M is a different file, 0.17 GB larger than Unsloth's used here. No per-level quality loss has been published.
131K
Max Context
2
Quant Variants
GGUF Q8_0
Best Quality
—
Perplexity rise (best build)
Quantization Variants
Per-quant VRAM, quality loss, and inference speed on RTX 4090
Measured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports
| Format | Level | BPW | VRAM | PPL Loss | Speed | Source | Actions |
|---|---|---|---|---|---|---|---|
| GGUF | Q4_K_M | 4.89 | 14.2 GB | — | — | Estimated | |
| GGUF | Q8_0 | 8.5 | 24.5 GB | — | — | Estimated |
Which file each size is for
A file size is not a memory figure: the VRAM column is the calculator's total at 4K context — weights plus KV cache plus a 10% activation buffer. Units differ on purpose: file sizes are decimal GB, as Hugging Face lists them; the VRAM column is GiB (1024³ bytes) with 2% added to the weights, so 7.12 GB of file counts as about 6.8 GiB there. A vision projector (mmproj), where the repo ships one, is a separate download and is in neither number.
- GGUF Q4_K_Munsloth/ERNIE-4.5-21B-A3B-PT-GGUF · ERNIE-4.5-21B-A3B-PT-Q4_K_M.gguf · 13.33 GB (13,331,015,456 bytes) · listing read 2026-10-21
- GGUF Q8_0bartowski/baidu_ERNIE-4.5-21B-A3B-PT-GGUF · baidu_ERNIE-4.5-21B-A3B-PT-Q8_0.gguf · 23.21 GB (23,205,354,592 bytes) · listing read 2026-10-21
Related community runs
These runs describe their own settings. Their speeds do not replace this estimate or prove your configuration fits.
Our reviewed link index has no run for this model yet. You can browse community reports or share your own test below.
Browse more runs on llamaPerf ↗Similar models
Compare with Qwen3 30B-A3BQwen3 30B-A3B Instruct
Alibaba Qwen3
Qwen3 MoE with only 3B active params. Q4 ~19GB file; outperforms QwQ-32B on 16GB cards.
GPT-OSS 20B
OpenAI GPT-OSS
OpenAI open-weight MoE (21B total / 3.6B active), shipped natively in MXFP4 — ~12.8GB runs on a 16GB card with no quality tax. Only 3.6B active params means CPU-offload stays usable.
Qwen3-VL 30B-A3B Instruct
Alibaba Qwen3-VL
Multimodal MoE with only ~3B active parameters, so it stays responsive on Apple unified memory and survives CPU offload far better than a dense 30B. At Q4_K_M it needs about 19.9 GB at 4K, comfortable on a 24 GB card; Q3_K_M brings that to 15.1 GB. The vision encoder is a separate 1.1 GB file.
Magistral Small 1.2 24B
Mistral Magistral
Mistral's reasoning model on a Mistral Small 3.2 base, with reasoning wrapped in [THINK] tags. At Q4_K_M it needs about 15.7 GB at 4K, tight on a 16 GB card and comfortable on 20 GB — explicit chain-of-thought a level below the 70B-class reasoners most guides assume.
Running ERNIE 4.5 21B-A3B locally
At Q4_K_M and 4K of context, ERNIE 4.5 21B-A3B needs about 14.2 GB — 12.7 GB of weights, 0.22 GB of KV cache and a 1.3 GB activation buffer. The smallest card in this index that clears that comfortably is the Radeon RX 7900 XT at 20 GB, and 34 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs.
What longer context costs
Going from 4K to 32K adds about 1.53 GB, taking the total to 15.9 GB. Weights do not move with context — only the KV cache does, and it grows linearly, so this is the number to watch when planning for long documents. The model's native window is 128K; holding all of it at Q4_K_M would need about 22 GB.
Which build to download
This index only tracks ERNIE 4.5 21B-A3B in GGUF, across 2 levels (Q4_K_M, Q8_0). No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving.
Common questions
- How much VRAM does ERNIE 4.5 21B-A3B need?
- About 14.2 GB at Q4_K_M with 4K of context and batch 1, rising to roughly 15.9 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card.
- Will ERNIE 4.5 21B-A3B run on a 20GB GPU?
- Yes — at Q4_K_M and 4K context it needs about 14.2 GB, which leaves 5.8 GB spare on a Radeon RX 7900 XT. That is the smallest card in this index that clears it comfortably; 34 of 78 do.
- Which quantization of ERNIE 4.5 21B-A3B should I use?
- No published quality comparison exists for this model, so pick by footprint: Q4_K_M is the level most people run, and a higher bits-per-weight level is more faithful. All 2 levels in the index are Q4_K_M, Q8_0. Note this is a mixture-of-experts model — all parameters must be resident even though only a fraction are active per token, so the memory cost follows the total, not the active count.
Where this model fits
Sized at Q4_K_M with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory.
- Runs comfortably on
- Radeon RX 7900 XT20GB · +5.8Mac M4 24G24GB · +3.8Mac M4 Pro 24G24GB · +3.8RTX 309024GB · +9.8Radeon RX 7900 XTX24GB · +9.8
- Tight but possible
- RTX 4060 Ti 16G16GBRadeon RX 7600 XT16GBRadeon RX 9060 XT 16G16GB
- Ships in
- GGUF
- Try it yourself
- Size it in the calculatorGenerate the run command
How to actually run this
Deployment guides for this model and this class of hardware.
This page's figures change when the model or the runtime does.Last updated 2026-11-17 RSS → /feed.xml