Alibaba (Qwen) · Apache 2.0
Qwen3.8 27B
A leading 2026 all-round model that fits a 24 GB graphics card. Reads images and video too.
How much memory it needs
The model file, plus memory for the conversation, plus about 1 GB for the app. Longer conversations need more.
| Size | Version | File | At 8K | At 32K | At 128K |
|---|---|---|---|---|---|
| 28B | UD-IQ3_S | 12.0 GB | 14.5 GB | 16.1 GB | 22.5 GB |
| UD-IQ4_XS | 14.3 GB | 16.8 GB | 18.4 GB | 24.8 GB | |
| Q4_K_M | 17.4 GB | 19.9 GB | 21.5 GB | 28.0 GB | |
| Q8_0 | 29.1 GB | 31.6 GB | 33.2 GB | 39.6 GB |
Which computers run it
The best size for each machine at an 8K-token conversation. Speeds are estimates from memory bandwidth.
| Computer | Best version | Fit | Speed |
|---|---|---|---|
| Laptop or PC, no graphics card, 32 GB | 28B Q4_K_M | Runs well | 3–5tok/sSlow |
| RTX 4060 · 8 GB | 28B UD-IQ3_S | Spills into system RAM | 6.5–10tok/sReadable |
| RTX 3060 · 12 GB | 28B UD-IQ3_S | Spills into system RAM | 11–17tok/sReadable |
| RTX 5060 Ti · 16 GB | 28B UD-IQ3_S | Runs well | 20–35tok/sSmooth |
| RTX 4090 · 24 GB | 28B Q4_K_M | Runs well | 35–50tok/sFast |
| RTX 5090 · 32 GB | 28B Q4_K_M | Runs great | 55–90tok/sFast |
| Mac with M4 · 16 GB | 28B UD-IQ3_S | Fits with a tweak | 5.5–9.5tok/sReadable |
| Mac with M4 Pro · 48 GB | 28B Q8_0 | Runs well | 5.5–9tok/sReadable |
| Mac with M5 Max · 128 GB | 28B Q8_0 | Runs great | 12–20tok/sSmooth |
| Ryzen AI Max+ 395 · 128 GB | 28B Q4_K_M | Runs great | 8.5–14tok/sReadable |
| DGX Spark · 128 GB | 28B Q8_0 | Runs great | 5.5–9tok/sReadable |
| RTX PRO 6000 · 96 GB | 28B Q8_0 | Runs great | 35–55tok/sFast |
| Mac Studio with M5 Ultra · 256 GB | 28B Q8_0 | Runs great | 18–30tok/sSmooth |
| Mac Studio with M3 Ultra · 512 GB | 28B Q8_0 | Runs great | 12–20tok/sSmooth |
Measured speeds
Real runs, not estimates. All measurements.
| Computer | Version | Prompt | Speed | Source |
|---|---|---|---|---|
| MacBook Pro (M5 Pro, 64 GB) | UD-Q4_K_XL | 512 | 13.4 tok/s | Ours |
| MacBook Pro (M5 Pro, 64 GB) | UD-Q4_K_XL | 32K | 10.28 tok/s | Ours |
What it is good and bad at
Good at
- Coding and multi-step agent work
- Reading images, screenshots and video
- Long conversations: its conversation memory is small for its size
Not good at
- Needs about 18 GB even at Q4 — too big for 16 GB cards
- Dense model, so slower per word than mixture-of-experts models
Start it
Install Ollama, then run one of these in a terminal. LM Studio users can search the same name in the app.
ollama run qwen3.8:27b