How Much VRAM Does Qwen3.8 2.4T A95B Need to Run Locally?
See devices that run itQwen3.8 2.4T A95B is a MoE model with 2,446B total parameters and 95B active per token. At Q4 it needs 1,366GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 30GB of VRAM plus 1,337GB of RAM. No single card in the database fits it; it needs multiple GPUs or the unified memory of a Mac Studio. In offload mode the cheapest card that runs it is the Arc Pro B65, at roughly 2 tokens/s.
Qwen3.8 2.4T A95B details
VRAM needed for Qwen3.8 2.4T A95B by quantization and context
Total at 8KMoE: all 2,446B parameters must stay resident in memory, while speed is set by the 95B active per token.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 564GGUF | 575.8 | 577.9 | 586.6 |
| 2-bit | 730.7GGUF | 910.6 | 912.8 | 921.4 |
| 3-bit | 955.5GGUF | 1,173.8 | 1,175.9 | 1,184.5 |
| 4-bit | 1,310.9GGUF | 1,365.1 | 1,367.3 | 1,375.9 |
| 5-bit | 1,650.4est. | 1,652.1 | 1,654.3 | 1,662.9 |
| 8-bit | 2,600.2GGUF | 2,537.1 | 2,539.3 | 2,547.9 |
| FP16 | 4,893.2GGUF | 4,785.6 | 4,787.7 | 4,796.3 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 256K limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.
Which GPUs can run Qwen3.8 2.4T A95B?
Q4_K_M · 8K context · sorted by eBay used priceNo single card in the database fits Qwen3.8 2.4T A95B: it needs at least 1,366GB of VRAM at Q4_K_M with an 8K context.
A multi-GPU build needs at least 1,366GB of VRAM in total.
None of the unified-memory machines in the database fit it either.
GPUs that run Qwen3.8 2.4T A95B with experts offloaded to RAM
Expert weights live in system RAM (needs at least 1,337 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 30 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.
| Device | VRAMGB | Used price | RAM neededGB | Est. t/s |
|---|---|---|---|---|
 Arc Pro B65 | 32 | — | 1,337 | 1.5 |
| Arc Pro B70 | 32 | — | 1,337 | 1.5 |
 GeForce RTX 4080 SUPER 32GB (modded) | 32 | — | 1,337 | 1.7 |
| GeForce RTX 5090 | 32 | — | 1,337 | 1.8 |
| GeForce RTX 5090 D | 32 | — | 1,337 | 1.8 |
 Instinct MI100 | 32 | — | 1,337 | 1.7 |
| Radeon AI PRO R9700 | 32 | — | 1,337 | 1.6 |
| Radeon PRO W7800 | 32 | — | 1,337 | 1.5 |
| RTX PRO 4500 Blackwell | 32 | — | 1,337 | 1.7 |
| Tesla V100 32GB | 32 | — | 1,337 | 1.7 |
| A100 40GB PCIe | 40 | — | 1,337 | 1.7 |
| CMP 170HX 40GB (modded) | 40 | — | 1,337 | 1.7 |
| GeForce RTX 4090 48GB (modded) | 48 | — | 1,337 | 1.7 |
| Radeon PRO W7900 | 48 | — | 1,337 | 1.6 |
| RTX PRO 5000 Blackwell 48GB | 48 | — | 1,337 | 1.7 |
| CMP 170HX 64GB (modded) | 64 | — | 1,337 | 1.7 |
| Instinct MI210 | 64 | — | 1,337 | 1.7 |
| RTX PRO 5000 Blackwell 72GB | 72 | — | 1,337 | 1.7 |
| RTX PRO 6000D | 84 | — | 1,337 | 1.7 |
| RTX PRO 6000 Blackwell | 96 | — | 1,337 | 1.8 |
| DGX Spark 128GB | 128 | — | 1,337 | 1.4 |
| Mac Studio M4 Max 128GB | 128 | — | 1,337 | 1.6 |
| Mac Studio M5 Max 128GB | 128 | — | 1,337 | 1.6 |
| MacBook Pro M4 Max 128GB | 128 | — | 1,337 | 1.6 |
| MacBook Pro M5 Max 128GB | 128 | — | 1,337 | 1.6 |
| Ryzen AI Max+ 395 128GB | 128 | — | 1,337 | 1.3 |
| Mac Studio M3 Ultra 512GB | 512 | — | 1,337 | 1.6 |
Run Qwen3.8 2.4T A95B with Ollama, llama.cpp or vLLM
The Ollama library has no official tag for it yet.
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf unsloth/Qwen3.8-2.4T-A95B-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve Qwen/Qwen3.8-2.4T-A95B
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Qwen3.8 2.4T A95B need?
At Q4_K_M with an 8K context Qwen3.8 2.4T A95B needs about 1,366GB of VRAM; 1,640GB leaves comfortable headroom.
What is the cheapest GPU that runs Qwen3.8 2.4T A95B?
No single card in the database fits Qwen3.8 2.4T A95B; it needs multiple GPUs or a Mac Studio's unified memory.
Can you run Qwen3.8 2.4T A95B with Ollama?
Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.