Nanbeige4.3 33B: VRAM requirements & speed
Nanbeige4.3 33B (33.12B dense, apache-2.0). Native context 512k. Smallest GGUF: 20.79 GB.
Memory needed by quantization and context
GGUF file size plus f16 KV cache (and fixed overhead of ~1 GB per GPU not shown). Source: bartowski/Nanbeige_Nanbeige4.3-33B-GGUF.
| Quant | File | +8k ctx | +32k ctx | +128k ctx | +256k ctx |
|---|---|---|---|---|---|
| Q4_K_M | 20.79 GB | 22.5 GB | 27.8 GB | 48.7 GB | 76.6 GB |
| Q5_K_M | 25.07 GB | 26.8 GB | 32.0 GB | 53.0 GB | 80.9 GB |
| Q6_K | 29.02 GB | 30.8 GB | 36.0 GB | 56.9 GB | 84.9 GB |
| Q8_0 | 35.2 GB | 36.9 GB | 42.2 GB | 63.1 GB | 91.0 GB |
KV cache size by context length
Computed from the model's attention layout.
| Context | f16 | q8_0 | q4_0 |
|---|---|---|---|
| 8k | 1.74 GB | 0.93 GB | 0.49 GB |
| 32k | 6.98 GB | 3.71 GB | 1.96 GB |
| 128k | 27.92 GB | 14.83 GB | 7.85 GB |
| 256k | 55.83 GB | 29.66 GB | 15.70 GB |
Which hardware runs Nanbeige4.3 33B? · 16k context
| Hardware | Quant | Fit | Memory | Tok/s | Speed source |
|---|---|---|---|---|---|
| AMD Ryzen AI Max+ 395 (Strix Halo) 128GB | Q4_K_M | fits fully | 24.3 GB | 75 | scaled from measured |
| RTX 5090 32GB | Q5_K_M | fits fully | 28.6 GB | 64 | scaled from measured |
| RTX PRO 6000 96GB | Q8_0 | fits fully | 38.7 GB | 41 | scaled from measured |
| Mac Studio M5 Ultra 256GB | Q8_0 | fits fully | 38.7 GB | 21 | estimated |
| Mac Studio M5 Ultra 512GB (ships late Oct 2026) | Q8_0 | fits fully | 38.7 GB | 21 | estimated |
| RTX 4090 24GB | Q4_K_M | partial offload (slow) | 24.3 GB | 20 | estimated |
| RTX 3090 24GB | Q4_K_M | partial offload (slow) | 24.3 GB | 19 | estimated |
| 2x RTX 3090 (48GB total) | Q8_0 | fits fully | 38.7 GB | 18 | estimated |
| AMD RX 7900 XTX 24GB | Q4_K_M | partial offload (slow) | 24.3 GB | 18 | estimated |
| Mac Studio M5 Max 128GB | Q6_K | fits fully | 32.5 GB | 17 | scaled from measured |
| MacBook Pro 14/16 M5 Max 128GB | Q6_K | fits fully | 32.5 GB | 17 | scaled from measured |
| Mac Studio (2025) M3 Ultra 96GB | Q4_K_M | fits fully | 24.3 GB | 16 | estimated |
| MacBook Pro 14/16 (2024) M4 Max 128GB | Q4_K_M | fits fully | 24.3 GB | 12 | estimated |
| NVIDIA RTX Spark N1X 128GB | Q4_K_M | fits fully | 24.3 GB | 12 | estimated |
| NVIDIA DGX Spark 128GB | Q4_K_M | fits fully | 24.3 GB | 11 | estimated |
| AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GB | Q4_K_M | fits fully | 24.3 GB | 10 | estimated |
| MacBook Pro 14/16 M5 Pro 48GB | Q4_K_M | fits fully | 24.3 GB | 10 | estimated |
| Mac mini (2024) M4 Pro 64GB | Q4_K_M | fits fully | 24.3 GB | 8 | estimated |
| RTX 5080 16GB | Q4_K_M | partial offload (slow) | 24.3 GB | 6 | estimated |
| RTX 5070 Ti 16GB | Q4_K_M | partial offload (slow) | 24.3 GB | 6 | estimated |
| RTX 5060 Ti 16GB | Q4_K_M | partial offload (slow) | 24.3 GB | 6 | estimated |
| 16GB VRAM class (RTX 4060 Ti 16GB, 32GB RAM) | Q4_K_M | partial offload (slow) | 24.3 GB | 5 | estimated |
| RTX 5070 12GB | Q4_K_M | partial offload (slow) | 24.3 GB | 4 | estimated |
| RTX 4070 12GB | Q4_K_M | partial offload (slow) | 24.3 GB | 4 | estimated |
| 8GB VRAM class (RTX 4060 8GB, 32GB RAM) | Q4_K_M | partial offload (slow) | 24.3 GB | 3 | estimated |
| RTX 3060 12GB (32GB RAM) | Q4_K_M | partial offload (slow) | 24.3 GB | 2 | estimated |
| Mac mini / iMac M4 16GB | - | does not fit | - | - | |
| Mac mini M6 32GB | - | does not fit | - | - | |
| MacBook Air M5 32GB | - | does not fit | - | - |
Discrete GPUs are modelled with 32 GB of system RAM available for offload. Does not fit even with offload: Mac mini / iMac M4 16GB, Mac mini M6 32GB, MacBook Air M5 32GB.
MLX conversions for Apple Silicon
No MLX conversion yet: we found no verified mlx-community 4-bit or 8-bit repo for Nanbeige4.3 33B. Use the GGUF files with llama.cpp or ollama on a Mac. How we check.
Links
Model card · GGUF files · Open in the advisor · last verified 2026-10-26
Can it run? Nanbeige4.3 33B on: RTX 5090 32GB (all pairs)
FAQ
How much VRAM does Nanbeige4.3 33B need?
The smallest quant (Q4_K_M) is 20.79 GB, so you need roughly 24 GB for an 8k context including KV cache and overhead. At 256k context the f16 KV cache alone is 55.8 GB.
Can I run Nanbeige4.3 33B on a AMD Ryzen AI Max+ 395 (Strix Halo) 128GB?
Yes: on AMD Ryzen AI Max+ 395 (Strix Halo) 128GB it fits fully as Q4_K_M at about 75 tok/s (scaled from measured). 15 of 29 hardware configurations fit it fully at 16k context.
What is the maximum context for Nanbeige4.3 33B?
512k tokens natively. Memory, not the model, is usually the limit: see the KV cache table above.
All models: model index · by memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB. Numbers use the same fit math as the advisor; sources and methods.