Qwen3-Next 80B-A3B VRAM Requirements
Qwen3-Next 80B-A3B needs about 46.3 GB of VRAM at Q4_K_M with an 8K context: 45.2 GB for weights, 0.3 GB for the KV cache and 0.8 GB of runtime overhead. At Q8_0 it needs 80.2 GB and at full FP16 precision 150 GB. The smallest GPU that runs it fully is the Apple M5 Pro 64GB.
46.3 GB VRAM
Qwen3-Next 80B-A3B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 45.2 GB
- KV cache 0.3 GB
- Runtime overhead 0.8 GB
Mixture of experts: all 80B parameters need memory, but only 3B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GB33% on GPU~25 tok/s
- RTX 3090 24GB49% on GPU~32 tok/s
- Radeon RX 7900 XTX 24GB49% on GPU~32 tok/s
- RTX 4090 24GB49% on GPU~33 tok/s
- RTX 5090 32GB65% on GPU~44 tok/s
- RTX 6000 Ada 48GBTight fit~125 tok/s
- A100 80GBFits~159 tok/s
- H100 80GB SXMFits~175 tok/s
- Ryzen AI Max+ 395 128GBFits~56 tok/s
- Apple M4 Max 128GBFits~96 tok/s
- Apple M5 Max 128GBFits~102 tok/s
- RTX PRO 6000 Blackwell 96GBFits~153 tok/s
- NVIDIA DGX Spark 128GBFits~59 tok/s
- Apple M3 Ultra 512GBFits~117 tok/s
Multi-GPU: about 3× 24 GB cards (RTX 3090 / 4090) with the model split across them.
Qwen3-Next 80B-A3B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 45.2 GB | 46.3 GB | 46.8 GB | 64 GB |
| Q5_K_M | 5.69 | 53.0 GB | 54.1 GB | 54.6 GB | 64 GB |
| Q6_K | 6.59 | 61.4 GB | 62.5 GB | 63.0 GB | 80 GB |
| Q8_0 | 8.50 | 79.2 GB | 80.2 GB | 80.8 GB | 96 GB |
| FP16 / BF16 | 16.00 | 149 GB | 150 GB | 151 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 36.4 GB | 37.5 GB | 38.1 GB | 48 GB |
| Q2_K | 3.00 | 27.9 GB | 29.0 GB | 29.6 GB | 32 GB |
VRAM by context length
At Q4_K_M. Hybrid attention: only 12 of 48 layers use full attention and keep a KV cache (2 KV heads × 256 dimensions). The other 36 use linear attention (Gated DeltaNet) with a fixed state of about 75 MB that does not grow with context.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.2 GB | 0.1 GB | 46.2 GB | 46.1 GB |
| 8K tokens | 0.3 GB | 0.2 GB | 46.3 GB | 46.2 GB |
| 32K tokens | 0.8 GB | 0.5 GB | 46.8 GB | 46.5 GB |
| 128K tokens | 3.1 GB | 1.7 GB | 49.1 GB | 47.7 GB |
| 256K tokens | 6.1 GB | 3.3 GB | 52.1 GB | 49.3 GB |
GPUs that can run Qwen3-Next 80B-A3B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Apple M5 Pro 64GB | 48.0 GB | Tight fit | IQ4_XS | – | ~68 tok/s |
| Apple M1 Max 64GB | 48.0 GB | Tight fit | IQ4_XS | – | ~80 tok/s |
| Apple M4 Max 64GB | 48.0 GB | Tight fit | IQ4_XS | – | ~96 tok/s |
| RTX A6000 48GB | 48.0 GB | Tight fit | IQ4_XS | – | ~114 tok/s |
| L40S 48GB | 48.0 GB | Tight fit | IQ4_XS | – | ~120 tok/s |
| RTX 6000 Ada 48GB | 48.0 GB | Tight fit | IQ4_XS | – | ~125 tok/s |
| A100 80GB | 80.0 GB | Fits | Q6_K | 256K | ~159 tok/s |
| H100 80GB SXM | 80.0 GB | Fits | Q6_K | 256K | ~175 tok/s |
| Ryzen AI Max+ 395 128GB | 96.0 GB | Fits | Q8_0 | 256K | ~56 tok/s |
| Apple M3 Max 128GB | 96.0 GB | Fits | Q8_0 | 256K | ~80 tok/s |
| Apple M4 Max 128GB | 96.0 GB | Fits | Q8_0 | 256K | ~96 tok/s |
| Apple M5 Max 128GB | 96.0 GB | Fits | Q8_0 | 256K | ~102 tok/s |
| RTX PRO 6000 Blackwell 96GB | 96.0 GB | Fits | Q8_0 | 256K | ~153 tok/s |
| NVIDIA DGX Spark 128GB | 112 GB | Fits | Q8_0 | 256K | ~59 tok/s |
| A100 40GB | 40.0 GB | 81% on GPU | Q3_K_M | – | ~65 tok/s |
| RTX 5090 32GB | 32.0 GB | 65% on GPU | Q2_K | – | ~44 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | 65% on GPU | Q2_K | – | ~40 tok/s |
| RTX 3090 24GB | 24.0 GB | 49% on GPU | – | – | ~32 tok/s |
| RTX 3090 Ti 24GB | 24.0 GB | 49% on GPU | – | – | ~33 tok/s |
| RTX 4090 24GB | 24.0 GB | 49% on GPU | – | – | ~33 tok/s |
Multi-GPU: about 3× 24 GB cards (RTX 3090 or 4090) run it at Q4_K_M, using tensor or layer splitting in llama.cpp, vLLM or ExLlamaV2.
About Qwen3-Next 80B-A3B
The experimental model that introduced Qwen's hybrid Gated DeltaNet plus gated attention design, later used across Qwen 3.5. It has 512 experts but only about 3B active parameters, so it is very fast once loaded. The full 80B must sit in memory, which makes it a good fit for 64 GB Macs and 48 GB GPU setups.
| Spec | Value |
|---|---|
| Parameters | 80B total, 3B active (MoE) |
| Layers | 48 |
| Hidden size | 2048 |
| Attention heads / KV heads | 16 / 2 |
| Head dimension | 256 |
| Vocabulary | 151,936 |
| Max context (used here) | 256K tokens |
| Linear-attention layers | 36 of 48 (Gated DeltaNet, fixed 75 MB state) |
| KV cache per token (FP16) | 24 KB |
| Released | 2025-09 |
| License | Apache 2.0 |
Qwen3-Next 80B-A3B VRAM: common questions
Is Qwen3-Next 80B-A3B good for a 64GB Mac?
Yes. Just. Q4_K_M needs about 46 GB, right at the roughly 48 GB a 64 GB Mac gives the GPU by default, so keep the context short or raise the wired-memory limit. Once loaded it generates at small-model speed.
How much VRAM does Qwen3-Next 80B-A3B need?
About 46.3 GB at Q4_K_M with an 8K context, 80.2 GB at Q8_0 and 150 GB at FP16. Longer contexts add 0.02 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Qwen3-Next 80B-A3B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the Apple M5 Pro 64GB (48.0 GB usable). With 24 GB cards you need about 3 of them.
Does Qwen3-Next 80B-A3B need VRAM for all 80B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 3B parameters are read per token, so it generates text about as fast as a 3B dense model.