Can I run Qwen3-Next 80B-A3B?
Qwen3-Next 80B-A3B by Alibaba needs around 64 GB of RAM at the recommended 4-bit quantization (48.5 GB download). Your hardware is checked below — instantly, nothing leaves your browser. Expect roughly ~192 tok/s on a Apple M-series Max.
Reading your hardware signals…
Real-world notes
Qwen3-Next 80B-A3B is a mixture-of-experts model from Alibaba, and the headline number is misleading in a useful way: it has 80B parameters total but only routes about 3B of them per token. That means it generates at the speed of a tiny model while still drawing on the knowledge of a large one. The catch is memory. You still have to hold the whole thing in RAM or VRAM, so even at a 4-bit quant around 48.5 GB you need at least 64 GB to run it, which rules out single consumer GPUs and points you at a big-memory machine instead.
In practice this is an Apple Silicon or workstation model, not a gaming-GPU one. On an M Max with enough unified memory it streams at roughly 192 tokens per second, which is genuinely quick for an 80B-class model and the main reason people reach for it. A 12 GB RTX 3060 or a 24 GB RTX 4090 simply cannot load it. If you are on CPU with DDR5 you can still run it at about 28 tokens per second, slower but workable for batch jobs. The 256K context is real, but it is hungry: pushing toward 128K can take total memory to around 95.8 GB, so plan headroom before you fill it.
Against Llama 3.1 70B, the obvious dense comparison, Qwen3-Next tends to feel markedly faster for its size because only 3B parameters are active each step, where the Llama has to run all 70B. Quality on chat and reasoning is competitive, though dense 70B models can still hold an edge on the hardest multi-step prompts. Its standout trait is that speed-to-size ratio: large-model breadth at small-model latency, if you have the memory to host it. License is Apache 2.0, so you are free to use it commercially and in production without provider-specific strings.
Specifications
Size by quantization
| Quantization | Bits/weight | Download | Min RAM | Quality |
|---|---|---|---|---|
| Q2_K | 3.35 | 33.5 GB | 48 GB | Noticeable loss |
| Q4_K_MRecommended | 4.85 | 48.5 GB | 64 GB | Recommended |
| Q5_K_M | 5.65 | 56.5 GB | 96 GB | High |
| Q8_0 | 8.5 | 85.0 GB | 128 GB | Near-original |
| F16 | 16 | 160.0 GB | 256 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 | ~1.5 GB | ~50.0 GB |
| 8K tokens | ~3.0 GB | ~51.5 GB |
| 32K tokens | ~11.8 GB | ~60.3 GB |
| 128K tokens | ~47.3 GB | ~95.8 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 | Won't fit in VRAM |
| Apple M-series (base) | 100 GB/s | ~47 tok/s |
| Apple M-series Pro | 270 GB/s | ~126 tok/s |
| Apple M-series Max | 410 GB/s | ~192 tok/s |
| CPU only (dual-channel DDR5) | 60 GB/s | ~28 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.
Run it locally
The easiest path is Ollama — one command and you're chatting:
ollama run qwen3-next:80b