Can I run EXAONE 4.5 33B?
EXAONE 4.5 33B by LG AI Research needs around 32 GB of RAM at the recommended 4-bit quantization (20.0 GB download). Your hardware is checked below — instantly, nothing leaves your browser. Expect roughly ~17 tok/s on a Apple M-series Max.
Reading your hardware signals…
Real-world notes
EXAONE 4.5 33B is LG AI Research's vision-and-reasoning model, and at 33B dense parameters it sits firmly in the heavyweight tier of what you can self-host. A 4-bit quant lands around 20 GB, so it will not fit on a 12 GB card like an RTX 3060 at all; you need a 24 GB GPU such as a 4090, or a Mac with plenty of unified memory. The minimum to load it sensibly is about 32 GB of RAM. This is not a casual laptop model. It is for people who want a capable multimodal assistant and have the hardware to back it.
In daily use on a 4090 it runs around 43 tokens per second at 4-bit, which is comfortable for chat and reading images, though noticeably heavier than a small 8B model. On an Apple M Max you are looking at roughly 17 tokens per second, usable but more deliberate, and on CPU with DDR5 it drops to about 3 tokens per second, which is patience-testing. The 256K context window is large, but memory is the catch: even at 128K context the total footprint climbs to about 51.7 GB, so a 24 GB card forces you to keep working context modest.
Against Qwen3-VL 32B, the closest comparison in its class, EXAONE generally trades blows rather than dominating; Qwen3-VL tends to have broader tooling and quant availability, while Qwen 3 32B is the lighter pick if you only need text reasoning. EXAONE's standout trait is being a genuinely strong vision-plus-reasoning model in one package from a major lab. The important caveat is the license: it ships under the EXAONE License, which is non-commercial, so you cannot use it in production or any commercial product. Keep it to research, prototyping, and personal use.
Specifications
Size by quantization
| Quantization | Bits/weight | Download | Min RAM | Quality |
|---|---|---|---|---|
| Q2_K | 3.35 | 13.8 GB | 24 GB | Noticeable loss |
| Q4_K_MRecommended | 4.85 | 20.0 GB | 32 GB | Recommended |
| Q5_K_M | 5.65 | 23.3 GB | 32 GB | High |
| Q8_0 | 8.5 | 35.1 GB | 48 GB | Near-original |
| F16 | 16 | 66.0 GB | 96 GB | Original |
Sizes are estimates from parameter count × bits per weight; real GGUF builds vary slightly. · Data updated: 2026-06-11 · How we calculate these numbers →
Memory needed by context length
| Context | KV cache (est.) | Total memory (Q4) |
|---|---|---|
| 4K tokens | ~1.0 GB | ~21.0 GB |
| 8K tokens | ~2.0 GB | ~22.0 GB |
| 32K tokens | ~7.9 GB | ~27.9 GB |
| 128K tokens | ~31.7 GB | ~51.7 GB |
The KV cache grows with context length — a model that fits at 4K can run out of memory at 32K. Estimates assume an FP16 cache with grouped-query attention; actual usage varies by runtime.
Estimated speed by hardware
| Hardware | Bandwidth | ~Speed |
|---|---|---|
| NVIDIA RTX 3060 12GB | 360 GB/s | Won't fit in VRAM |
| NVIDIA RTX 4090 24GB | 1008 GB/s | ~43 tok/s |
| Apple M-series (base) | 100 GB/s | ~4 tok/s |
| Apple M-series Pro | 270 GB/s | ~11 tok/s |
| Apple M-series Max | 410 GB/s | ~17 tok/s |
| CPU only (dual-channel DDR5) | 60 GB/s | ~3 tok/s |
Token generation is memory-bandwidth bound: tok/s ≈ bandwidth × 0.85 ÷ model size at Q4. Real-world numbers vary by runtime and context length.