AlibabaQwen 3Qwen3NextForCausalLM

Qwen3 Next 80B A3B Thinking — Hardware Requirements & GPU Compatibility

ChatFunctions

Qwen3 Next 80B A3B Thinking is Alibaba's 81-billion-parameter mixture-of-experts model, with about 3 billion parameters active per token (the A3B in its name), drawn from 512 experts with only 10 routed at a time. Because only the active experts run per token, inference is far cheaper than a dense model this size, though the full weight set still needs to fit in memory. It combines mostly-linear Gated DeltaNet attention with periodic full attention layers, and this Thinking variant always reasons through a visible chain of thought before answering. Running it locally needs multiple GPUs or a large unified-memory machine. It supports a 262K token context window, is released under the Apache 2.0 license, and was published in September 2025 as part of the Qwen3-Next line introducing this hybrid-attention architecture.

81.3B params · MoE5.5B active262K contextApache 2.0Sep 202543.8K downloads503 likes47.3K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Decent on a 24 GB GPU

Runs on 24 GB (IQ2_XXS · ~200 tok/s · 64K)Macs from 32 GB (IQ2_M · ~61 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Qwen3 Next 80B A3B Thinking

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Your own private copy · Q4_K_M · 96 GB GPU

The easiest way to try this model
Running in a few minutes, easy setup
Chat right here or connect any app

$5+ bonus for new users

Run Qwen3 Next 80B A3B Thinking

49.3 GB VRAM at 16K context

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run qwen3-next:80b
Official Ollama library build. For the exact Q4_K_M file from unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF: ollama run hf.co/unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve Qwen/Qwen3-Next-80B-A3B-Thinking --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~179 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/Qwen3-Next-80B-A3B-Thinking-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
A100 80GBCheapestCheapest30 GB free	~194 tok/s	$1.63	Deploy on Vast.aifrom $1.14/hr (opens in new tab)	30 GB free · ~194 tok/s · $1.63 / 1M tokens
RTX PRO 6000 Max-Q (96 GB)46 GB free	~182 tok/s	$2.04	Deploy on Vast.aifrom $1.33/hr (opens in new tab)	96 GB · 46 GB free · ~182 tok/s · $2.04 / 1M tokens
A100 80GB30 GB free	~194 tok/s	$1.99	Deploy on RunPodfrom $1.39/hr (opens in new tab)	30 GB free · ~194 tok/s · $1.99 / 1M tokens
RTX PRO 6000 (96 GB)46 GB free	~182 tok/s	$2.35	Deploy on Vast.aifrom $1.54/hr (opens in new tab)	96 GB · 46 GB free · ~182 tok/s · $2.35 / 1M tokens
RTX PRO 6000 Server (96 GB)46 GB free	~171 tok/s	$2.75	Deploy on RunPodfrom $1.69/hr (opens in new tab)	96 GB · 46 GB free · ~171 tok/s · $2.75 / 1M tokens
RTX PRO 6000 Server (96 GB)46 GB free	~171 tok/s	$2.83	Deploy on Vast.aifrom $1.74/hr (opens in new tab)	96 GB · 46 GB free · ~171 tok/s · $2.83 / 1M tokens
H100 (80 GB)30 GB free	~240 tok/s	$3.12	Deploy on RunPodfrom $2.69/hr (opens in new tab)	80 GB · 30 GB free · ~240 tok/s · $3.12 / 1M tokens
H200 (141 GB)91 GB free	~269 tok/s	$3.70	Deploy on RunPodfrom $3.59/hr (opens in new tab)	141 GB · 91 GB free · ~269 tok/s · $3.70 / 1M tokens

Compare GPUs →

Your own private copy of Qwen3 Next 80B A3B Thinking · Q4_K_M · OpenAI-compatible API

The pod’s first start downloads ~49 GB, billed while it loads.

Open the pod on RunPod
Deploy on RunPodRTX PRO 6000 · $2.49/hr (opens in new tab)
New to RunPod? Sign up and add credit first. If that GPU is gone, pick another with at least 96 GB.
2
Set the model and your key
Press Set overrides → Environment variables. The first variable is already there; set its value:
Key
LLAMA_ARG_HF_REPO
Value
unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF:Q4_K_M
Press Add Environment variable for your key:
Key
LLAMA_API_KEY
Value
Make up your own key
A long password works. You’ll paste it below as the API key. Without it, anyone who finds the pod’s address can use it.Then Set overrides.
Deploy
Press Deploy Pod. Usually ready in 15–60 min.

Use it

RunPod → your pod → Connect → HTTP 8080, copy the link. The key is the one you made up in step 2.

Endpoint URL

API key

Chats and keys stay only in this browser.

https://<POD_ID>-8080.proxy.runpod.net/v1

RunPod → your pod → Connect → HTTP 8080 gives this address. The API key is the one from step 2.

from openai import OpenAI
client = OpenAI(base_url="https://<POD_ID>-8080.proxy.runpod.net/v1", api_key="YOUR_KEY")
r = client.chat.completions.create(
    model="any",
    messages=[{"role": "user", "content": "Hello!"}])
print(r.choices[0].message.content)

Works with any OpenAI client.

https://<POD_ID>-8080.proxy.runpod.net/v1

RunPod → your pod → Connect → HTTP 8080 gives this address. The API key is the one from step 2.

Free on Windows, macOS, Linux, iOS, Android and the web.

Open the web app ↗ (opens in a new tab)Download ↗ (opens in a new tab)

1. Settings → Model Provider → Add → Add Custom Provider
2. API Mode OpenAI API Compatible; paste the API key and the base URL as API Host
3. Model → Fetch, pick the model
4. New chat → pick the provider → start chatting

Free and open source for Windows, macOS and Linux.

Download Jan ↗ (opens in a new tab)

1.Settings → next to Model Providers press +
2.API format OpenAI-compatible; Base URL is the base URL above (with /v1); API key is your key; Create
3.If no model appears, click + in Models and type any
4.New chat → pick the model → start chatting

How Much VRAM Does Qwen3 Next 80B A3B Thinking Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
22.8 GB	23.1 GB	24.3 GB	26.2 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 64K context · ~200 tok/s
24.8 GB	25.1 GB	26.3 GB	22.2 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Does not fit: needs 24.9 GB
25.8 GB	26.1 GB	27.3 GB	23.4 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Does not fit: needs 25.9 GB
27.9 GB	28.2 GB	29.4 GB	26.1 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Does not fit: needs 28 GB
31.9 GB	32.2 GB	33.4 GB	33 GB	3-bit · i-quant — Importance-weighted 3-bit	Does not fit: needs 32 GB
32.9 GB	33.2 GB	34.4 GB	27.3 GB	2-bit · small K-quant — 2-bit small K-quant	Does not fit: needs 33 GB
34.0 GB	34.3 GB	35.5 GB	33 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Does not fit: needs 34.1 GB
35.0 GB	35.3 GB	36.5 GB	34.5 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Does not fit: needs 35.1 GB
35.0 GB	35.3 GB	36.5 GB	29.2 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Does not fit: needs 35.1 GB
36.0 GB	36.3 GB	37.5 GB	34.6 GB	3-bit small — 3-bit small quantization	Does not fit: needs 36.1 GB
37 GB	37.3 GB	38.5 GB	36.6 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Does not fit: needs 37.1 GB
40.0 GB	40.4 GB	41.6 GB	38.3 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 40.2 GB
41.1 GB	41.4 GB	42.6 GB	45.3 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 41.2 GB
42.1 GB	42.4 GB	43.6 GB	38.2 GB	3-bit large — 3-bit large quantization	Does not fit: needs 42.2 GB
44.1 GB	44.4 GB	45.6 GB	42.6 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Does not fit: needs 44.2 GB
46.1 GB	46.5 GB	47.7 GB	50.1 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Does not fit: needs 46.3 GB
46.1 GB	46.5 GB	47.7 GB	45.5 GB	4-bit small — 4-bit small quantization	Does not fit: needs 46.3 GB
46.1 GB	46.5 GB	47.7 GB	45.1 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Does not fit: needs 46.3 GB
49.2 GB	49.5 GB	50.7 GB	48.5 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 49.3 GB
50.2 GB	50.5 GB	51.7 GB	49 GB	4-bit large — 4-bit large quantization	Does not fit: needs 50.3 GB
56.3 GB	56.6 GB	57.8 GB	55 GB	5-bit small — 5-bit small quantization	Does not fit: needs 56.4 GB
58.3 GB	58.6 GB	59.9 GB	56.8 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 58.4 GB
59.4 GB	59.7 GB	60.9 GB	57.2 GB	5-bit large — 5-bit large quantization	Does not fit: needs 59.5 GB
67.5 GB	67.8 GB	69 GB	65.5 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 67.6 GB
81.7 GB	82 GB	83.2 GB	84.8 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 81.8 GB
163.1 GB	163.4 GB	164.6 GB	159.5 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 163.2 GB

Selected Q4_K_M

4-bit medium quantization

49.3GBat 16K context

Weights: 48.8 GB
KV cache: 0.2 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 25.3 GB more, IQ2_XXS fits

≥50 tok/s needs ≥296 GB/s · M3 Ultra ~65, M2 Ultra ~64

Or .

Which Hardware Can Run Qwen3 Next 80B A3B Thinking?

Q4_K_M · 49.2 GB

Qwen3 Next 80B A3B Thinking (Q4_K_M) requires 49.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 64+ GB is recommended. Using the full 262K context window can add up to 3.1 GB, bringing total usage to 52.3 GB. No consumer GPU has enough memory. 52 devices with unified memory can run Qwen3 Next 80B A3B Thinking, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 22.9 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceIQ2_XXS · 22.9 GB · ~200 tok/s · 64K22.9~20064K24 GB tier →

Mac 32 GB · Mac Mini M4IQ2_M · 28 GB · ~31 tok/s · 32K28~3132KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 49.3 GB · ~55 tok/s · 256K49.3~55256KMacs →

How good is Qwen3 Next 80B A3B Thinking?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

LMArena Text
LMArena · Oct 2026Arena
1369.2
#65/188 open
Arena

Qwen3 Next 80B A3B Thinking benchmark results

Alternatives to Qwen3 Next 80B A3B Thinking

Qwen 3 family · 79 models →

Similar models

5tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Kolibri 1 BF16Aleph-Alpha	—	22.5 GB	~237	same size
AliceAI Foundation 80B A3B Baseyandex	—	49.9 GB	—needs > 24 GB	same size
NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16NVIDIA	—	54.5 GB	—needs > 24 GB	same size
Hunyuan A13B InstructTencent	—	50.7 GB	—needs > 24 GB	same size
Kolibri 1Aleph-Alpha	—	47.9 GB	—needs > 24 GB	same size

Specifications

Publisher: Alibaba
Family: Qwen 3
Parameters: 81.3B
Active parameters: 5.5B per token
Architecture: Qwen3NextForCausalLM
Context Length: 262,144 tokens
Vocabulary Size: 151,936
Release Date: 2025-09-09
License: Apache 2.0

Resources

Frequently Asked Questions

How much VRAM does Qwen3 Next 80B A3B Thinking need?

Qwen3 Next 80B A3B Thinking requires 49.2 GB of VRAM at Q4_K_M, or 163.1 GB at BF16. Full 262K context adds up to 3.1 GB (52.3 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 81.3B × 4.8 bits ÷ 8 = 48.8 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 3.5 GB (at full 262K context)

VRAM usage by quantization

Q4_K_M

49.2 GB

Q4_K_M + full context

52.3 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Qwen3 Next 80B A3B Thinking?

Yes, at IQ2_XXS (22.8 GB) or lower. Higher quantizations like IQ2_XS (24.8 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Qwen3 Next 80B A3B Thinking?

For Qwen3 Next 80B A3B Thinking, Q4_K_M (49.2 GB) offers the best balance of quality and VRAM usage. Q4_K_L (50.2 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 22.8 GB.

VRAM requirement by quantization

IQ2_XXS

22.8 GB~53%

IQ3_XS

34.0 GB~73%

Q3_K_L

42.1 GB~86%

Q4_K_M ★

49.2 GB~89%

Q4_K_L

50.2 GB~90%

BF16

163.1 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Qwen3 Next 80B A3B Thinking on a Mac?

Yes — Mac Studio M1 Max (32 GB) and 84 other Macs can run Qwen3 Next 80B A3B Thinking, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 49.2 GB of usable unified memory (RAM minus macOS overhead), from Mac Pro M2 Ultra (64 GB).

Can I run Qwen3 Next 80B A3B Thinking locally?

Yes — Qwen3 Next 80B A3B Thinking can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Qwen3 Next 80B A3B Thinking?

At Q4_K_M, Qwen3 Next 80B A3B Thinking can reach ~97 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 3.3 GB active ÷ (8000 × 0.65) = 0.64 ms, plus 48 layers × 0.055 ms = 2.64 ms, so 1000 ÷ 3.28 ms = ~305 tok/s

Estimated speed at Q4_K_M (49.2 GB)

NVIDIA B200

~305 tok/s

NVIDIA B300

~305 tok/s

NVIDIA GH200 Grace Hopper Superchip

~271 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Qwen3 Next 80B A3B Thinking?

At Q4_K_M, the download is about 48.79 GB. The full-precision BF16 version is 162.65 GB. The smallest option (IQ2_XXS) is 22.36 GB.

Which GPUs can run Qwen3 Next 80B A3B Thinking?

No single consumer GPU has enough VRAM to run Qwen3 Next 80B A3B Thinking at Q4_K_M (49.2 GB). Multi-GPU or professional hardware is required.

Which devices can run Qwen3 Next 80B A3B Thinking?

52 devices with unified memory can run Qwen3 Next 80B A3B Thinking at Q4_K_M (49.2 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.