What hardware do I need?
Hardware that delivers the experience you want — not just loads the model
Pick the model and describe the workload. We rate every machine in the database for that exact workload and group them by whether they meet your target.
1 · Model
2 · Use case
Agentic coding
Chat & writing
Coding
Documents & research
Multimodal
Automation & serving
3 · Tool & runtime
Tool: OpenCode
4 · Workload
Project size, agent behaviour, other apps
5 · Desired experience
6 · Budget & vendors (optional)
Prices are approximate launch prices for the whole system.
Qwen3.8 27B · Agentic coding / medium repository
Target: comfortable · 16 machines meet it
Meets your comfortable target (16)
Cheapest first.
Desktop PC · GeForce RTX 3090 · 32 GB RAM
24 GB VRAM (936 GB/s) + 32 GB RAM
- Quant / runtime
- Q4 · Ollama
- Generation
- 30–40 tok/s
- Per step
- ≈20 s
- Headroom
- 4.3 GB
Desktop PC · Radeon RX 7900 XTX · 32 GB RAM
24 GB VRAM (960 GB/s) + 32 GB RAM
- Quant / runtime
- Q4 · Lemonade
- Generation
- 35–45 tok/s
- Per step
- ≈15 s
- Headroom
- 4.3 GB
Desktop PC · GeForce RTX 3090 · 64 GB RAM
24 GB VRAM (936 GB/s) + 64 GB RAM
- Quant / runtime
- Q4 · Ollama
- Generation
- 30–40 tok/s
- Per step
- ≈20 s
- Headroom
- 4.3 GB
Desktop PC · Radeon RX 7900 XTX · 64 GB RAM
24 GB VRAM (960 GB/s) + 64 GB RAM
- Quant / runtime
- Q4 · Lemonade
- Generation
- 35–45 tok/s
- Per step
- ≈15 s
- Headroom
- 4.3 GB
Desktop PC · Radeon AI PRO R9700 · 32 GB RAM
32 GB VRAM (640 GB/s) + 32 GB RAM
- Quant / runtime
- Q4 · Lemonade
- Generation
- 25–35 tok/s
- Per step
- ≈20 s
- Headroom
- 12 GB
Desktop PC · Radeon AI PRO R9700 · 64 GB RAM
32 GB VRAM (640 GB/s) + 64 GB RAM
- Quant / runtime
- Q4 · Lemonade
- Generation
- 25–35 tok/s
- Per step
- ≈20 s
- Headroom
- 12 GB
Desktop PC · GeForce RTX 4090 · 32 GB RAM
24 GB VRAM (1008 GB/s) + 32 GB RAM
- Quant / runtime
- Q5 · Ollama
- Generation
- 30–40 tok/s
- Per step
- ≈15 s
- Headroom
- 1.1 GB
Desktop PC · GeForce RTX 4090 · 64 GB RAM
24 GB VRAM (1008 GB/s) + 64 GB RAM
- Quant / runtime
- Q5 · Ollama
- Generation
- 30–40 tok/s
- Per step
- ≈15 s
- Headroom
- 1.1 GB
Desktop PC · 2× GeForce RTX 3090 (48 GB) · 128 GB RAM
48 GB VRAM (936 GB/s) + 128 GB RAM
- Quant / runtime
- Q5 · Ollama
- Generation
- 20–35 tok/s
- Per step
- ≈20 s
- Headroom
- 25 GB
Meets the acceptable target (32)
Usable, with noticeable compromises for this workload.
Mac Studio M2 Ultra 64 GB
64 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 33 GB
Mac Studio M2 Ultra 128 GB
128 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 97 GB
Mac Studio M2 Ultra 192 GB
192 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 161 GB
Mac Studio M3 Ultra 96 GB
96 GB unified memory · 819 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 65 GB
Mac Studio M3 Ultra 256 GB
256 GB unified memory · 819 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 225 GB
Mac Studio M3 Ultra 512 GB
512 GB unified memory · 819 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 481 GB
Mac Pro M2 Ultra 64 GB
64 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 33 GB
Mac Pro M2 Ultra 128 GB
128 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 97 GB
Mac Pro M2 Ultra 192 GB
192 GB unified memory · 800 GB/s
- Quant / runtime
- Q4 · LM Studio
- Generation
- 19–25 tok/s
- Per step
- ≈35 s
- Headroom
- 161 GB
Can run, but below your target (48)
The model loads, but the experience falls short for this workload.
Cannot run (32)
Not enough usable memory, or no compatible runtime.