unsloth
Qwen3 Next 80B A3B Instruct
Qwen3 Next 80B A3B Instruct is a 79.7B mixture-of-experts qwen3next model from unsloth, with about 3.0B parameters active per token. Configured context length is 262,144 tokens. License: apache-2.0. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters79.7B / 3.0B active
Context256K
Attentiongqa
Licenseapache-2.0
48 layers · 2 KV heads · head dim 256 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 148.5 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 86.7 GB | Hugging Face file bytes |
| Q8_0 | 79.0 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 63.8 GB | Hugging Face file bytes |
| Q6_K | 61.0 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 52.8 GB | Hugging Face file bytes |
| Q5_K_M | 52.9 GB | Hugging Face file bytes |
| Q5_K_S | 51.2 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 42.9 GB | Hugging Face file bytes |
| Q4_K_M | 45.2 GB | Hugging Face file bytes |
| Q4_K_S | 42.4 GB | Hugging Face file bytes |
| IQ4_NL | 42.0 GB | Hugging Face file bytes |
| IQ4_XS | 39.7 GB | Hugging Face file bytes |
| Q4_1 | 46.6 GB | Hugging Face file bytes |
| Q4_0 | 42.2 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 33.2 GB | Hugging Face file bytes |
| Q3_K_M | 35.7 GB | Hugging Face file bytes |
| Q3_K_S | 32.2 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 30.8 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 28.1 GB | Hugging Face file bytes |
| Q2_K | 54.4 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 24.4 GB | Hugging Face file bytes |
| UD-IQ1_M | 22.7 GB | Hugging Face file bytes |
| UD-IQ1_S | 21.3 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Expert offload | UD-Q2_K_XL | 82–111 |
| GeForce RTX 3060 12GB | Expert offload | UD-Q2_K_XL | 102–139 |
| GeForce RTX 4060 Ti 16GB | Expert offload | UD-Q2_K_XL | 82–111 |
| GeForce RTX 4090 | Expert offload | UD-Q2_K_XL | 286–390 |
| GeForce RTX 5090 | Tight fit | UD-Q2_K_XL | 641–874 |
| Apple M4 Pro | Won't fit | UD-IQ1_S | — |
| Apple M5 | Won't fit | UD-IQ1_S | — |
| CPU only · DDR5-5600 | Runs on CPU | UD-IQ1_M | 29–39 |
Smallest devices that fit it fully
- NVIDIA L4 — Tight fit, UD-IQ1_S, 130–178 tok/s
- NVIDIA A10 — Tight fit, UD-IQ1_S, 261–356 tok/s
- GeForce RTX 5090 — Tight fit, UD-Q2_K_XL, 641–874 tok/s
- RTX 5000 Ada — Tight fit, UD-Q2_K_XL, 206–281 tok/s
- Tesla V100 32GB — Tight fit, UD-Q2_K_XL, 255–348 tok/s
- Apple M1 Max — Tight fit, UD-IQ1_S, 132–180 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (UD-Q2_K_XL).
llama.cpp
# Expert offload is approximate. llama.cpp can pin experts to CPU: # -ot "\\.ffn_.*_exps\\.=CPU" llama-server -hf unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF:UD-Q2_K_XL -c 8192 -ngl 99
Ollama
ollama run qwen3
LM Studio
lms get unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF # Then load quant UD-Q2_K_XL