Qwen3 8B speed: tokens per second on GPUs and Macs

Qwen3 8B (ollama run qwen3:8b) is about 4.9 GB at Q4 (Ollama default). It fits fully on graphics cards with 8 GB of VRAM or more (38 of the 38 GPUs and Macs below); on smaller cards part of it runs on the CPU and slows down sharply. The fastest measured or published result below is 136 tokens/s on the Radeon RX 7900 XTX 24GB.

How fast does a model actually run on your GPU or Mac? Speeds here are measured by readers with Ollama and checked against the --verbose output they paste. Where readers haven't measured a combination yet, you see a published benchmark from a linked public source when one exists, otherwise our estimate from memory bandwidth, marked as an estimate.

Generation speed (Ollama "eval rate"), single user, default context. Results vary with drivers, context length and background load. Measured values are the median of all approved runs. As an Amazon Associate we earn from qualifying purchases.

Measure it yourself: ollama run qwen3:8b --verbose

Measured by readers (median)Published benchmark (linked source)Estimate, not measured yet
HardwareMemorySpeed (tokens/s)
Radeon RX 7900 XTX 24GB
Radeon RX 7900 XTX 24GB
Desktop GPU
24 GB
136
1 reader run estimate ~137
Check price ↗
Estimates: not measured on this model yet (bandwidth-based, usually optimistic)
RTX 5090 32GB
RTX 5090 32GB
Desktop GPU
32 GB
150+
Estimate (very fast)
Check price ↗
RTX 4090 24GB
RTX 4090 24GB
Desktop GPU
24 GB
~143
Estimate
Check price ↗
RTX 5080 16GB
RTX 5080 16GB
Desktop GPU
16 GB
~137
Estimate
Check price ↗
RTX 6000 Ada 48GB
RTX 6000 Ada 48GB
Desktop GPU
48 GB
~137
Estimate
Check price ↗
RTX 3090 24GB
Desktop GPU
24 GB
~133
Estimate
RTX 3080 12GB
Desktop GPU
12 GB
~130
Estimate
RTX 5070 Ti 16GB
RTX 5070 Ti 16GB
Desktop GPU
16 GB
~128
Estimate
Check price ↗
RTX 5090 Laptop 24GB
RTX 5090 Laptop 24GB
Laptop GPU · e.g. ASUS ROG Strix SCAR 18
24 GB
~128
Estimate
Check price ↗
RTX 5080 Laptop 16GB
RTX 5080 Laptop 16GB
Laptop GPU · e.g. MSI Vector A16
16 GB
~128
Estimate
Check price ↗
Radeon RX 7900 XT 20GB
Desktop GPU
20 GB
~114
Estimate
RTX 3080 10GB
Desktop GPU
10 GB
~108
Estimate
RTX 4080 / Super 16GB
Desktop GPU
16 GB
~102
Estimate
RTX 5070 12GB
RTX 5070 12GB
Desktop GPU
12 GB
~95.6
Estimate
Check price ↗
RTX 4070 Ti Super 16GB
RTX 4070 Ti Super 16GB
Desktop GPU
16 GB
~95.6
Estimate
Check price ↗
RTX 5070 Ti Laptop 12GB
Laptop GPU
12 GB
~95.6
Estimate
Radeon RX 9070 XT 16GB
Radeon RX 9070 XT 16GB
Desktop GPU
16 GB
~91.1
Estimate
Check price ↗
Radeon RX 7800 XT 16GB
Desktop GPU
16 GB
~88.8
Estimate
Mac Studio M3 Ultra 256GB
Mac Studio M3 Ultra 256GB
Mac
192 GB
~87.3
Estimate
Check price ↗
Mac Studio M2 Ultra 192GB
Mac
144 GB
~86.8
Estimate
RTX 4090 Laptop 16GB
Laptop GPU
16 GB
~82
Estimate
MacBook Pro 16" M5 Max 48GB
MacBook Pro 16" M5 Max 48GB
Mac
36 GB
~81.5
Estimate
Check price ↗
Mac Studio M4 Max 128GB
Mac Studio M4 Max 128GB
Mac
96 GB
~77.7
Estimate
Check price ↗
RTX 4070 12GB
Desktop GPU
12 GB
~71.7
Estimate
Mac Studio M5 Max 36GB
Mac Studio M5 Max 36GB
Mac
27 GB
~65.4
Estimate
Check price ↗
RTX 5060 Ti 16GB
RTX 5060 Ti 16GB
Desktop GPU
16 GB
~63.7
Estimate
Check price ↗
RTX 3070 8GB
Desktop GPU
8 GB
~63.7
Estimate
Mac M3 Max 64GB
Mac
48 GB
~56.9
Estimate
Mac M3 Max 128GB
Mac
96 GB
~56.9
Estimate
RTX 5060 Laptop 8GB
RTX 5060 Laptop 8GB
Laptop GPU · e.g. Lenovo Legion 5a
8 GB
~54.6
Estimate
Check price ↗
RTX 3060 12GB
Desktop GPU
12 GB
~51.2
Estimate
RTX 4060 Ti 16GB
RTX 4060 Ti 16GB
Desktop GPU
16 GB
~41
Estimate
Check price ↗
Mac M4 Pro 48GB
Mac
36 GB
~38.8
Estimate
RTX 4060 8GB
Desktop GPU
8 GB
~38.7
Estimate
RTX 4060 Laptop 8GB
RTX 4060 Laptop 8GB
Laptop GPU · e.g. Dell XPS 15
8 GB
~36.4
Estimate
Check price ↗
MacBook Air M5 16GB
MacBook Air M5 16GB
Mac
10.7 GB
~21.8
Estimate
Check price ↗
MacBook Pro 14" M5 16GB
MacBook Pro 14" M5 16GB
Mac
10.7 GB
~21.8
Estimate
Check price ↗
Mac mini M4 16GB
Mac mini M4 16GB
Mac
10.7 GB
~17.1
Estimate
Check price ↗

Add your result (2 minutes)

  1. Run the model with Ollama in verbose mode: ollama run qwen3-coder:30b --verbose
  2. Ask for a long answer, for example: Write a 300-word story about a robot learning to paint.
  3. Copy everything Ollama prints after the answer (from total duration to eval rate) and paste it below.

Results with the pasted Ollama output go live straight away when they are plausible; others are checked by hand first. We store an anonymous hash of your connection only to prevent duplicate submissions.

Latest measured runs

  • No measured runs yet. Be the first: add yours above.

Benchmarks by model

Speed tables for each model on every GPU, laptop GPU and Mac in our list.

All models: Local LLM Benchmarks overview

How these numbers work

What is measured?

Generation speed: how many tokens per second the model writes once it starts answering. That is Ollama's "eval rate" line. Prompt processing speed and load time are not included.

Where do the estimates come from?

From memory bandwidth: generation is limited by how fast the GPU or Mac can read the model's active weights, so we divide bandwidth by the model's size at that quantization. Estimates are shown only when the whole model fits in VRAM or unified memory; "doesn't fit" means part of it would run on the CPU and the speed depends heavily on your system.

What are "Published" results?

Real measurements from public benchmark projects, shown with a link to the source and its method in the tooltip. They are used only where the model, quantization and hardware match, and they step aside once readers have added three or more runs of their own. They may use llama.cpp directly rather than Ollama (Ollama is built on llama.cpp) and an older version, so treat small differences as normal.

Why a median?

The same card can differ by driver, power limit or laptop cooling. The median of several runs is a fair "typical" number, and one unusual result can't skew it.

Can I trust reader results?

Results go live automatically only with the pasted Ollama output, and only when they are close to other runs (or to the estimate when a combination is new). Everything else is reviewed by hand. The number of runs is shown next to every measured value.