Can a Strix Halo Framework Desktop, 128GB run local LLMs?
Yes — 33 of the 39 open models on this site fit in its 96 GB of usable memory, the strongest being Qwen3.8 27B. Whether that saves you money is a different question, and the answer is usually no.
Price$3,449
Memory128 GB, about 96 GB of it addressable by the GPU at 256 GB/s
Best model it runsQwen3.8 27B — Sonnet-class, 10 tok/s
Pay-back against the APIPays back in 45 years at 500k tokens a day
Run the numbers on this machine
What it runs
| Model | Speed | Class | Good at | Memory |
|---|---|---|---|---|
| Qwen3.8 27BQ4_K_M | 10 tok/s | Sonnet-class | 19 GB | |
| Ling 3.0 flashQ4_K_M | 20 tok/s | Haiku-class | 82 GB | |
| Qwen3.6 27BQ4_K_M | 10 tok/s | Haiku-class | 19 GB | |
| Qwen3.6 35B-A3BQ4_K_M | 56 tok/s | Haiku-class | 23 GB | |
| Muse Glimmer 30BQ4_K_M | 11 tok/s | Haiku-class | 18 GB | |
| Gemma 4 26B-A4BQ4_K_M | 41 tok/s | Haiku-class | 18 GB | |
| Qwen3.5 122B-A10BUD-Q4_K_M | 20 tok/s | Haiku-class | 79 GB | |
| Gemma 4 31B itQ4_K_M | 7.4 tok/s | Haiku-class | 26 GB | |
| Granite 4.2 30BQ4_K_M | 7.3 tok/s | Haiku-class | 26 GB | |
| GLM-4.7-FlashQ4_K_M | 40 tok/s | Haiku-class | 20 GB | |
| Gemma 4 12BQ4_K_M | 24 tok/s | Below every hosted tier | 8.0 GB | |
| Qwen3.5 9BQ4_K_M | 28 tok/s | Below every hosted tier | 6.8 GB |
21 more fit; the calculator lists them all.
The specifics
- Chip
- Framework Desktop — Ryzen AI Max+ 395 · Radeon 8060S, 40 CU
- Memory bandwidth
- 256 GB/s
- Usable by the GPU
- 96 GB — Windows lets you dedicate up to 75% of RAM to the GPU (Variable Graphics Memory in AMD Adrenalin; Framework documents 96 GB on a 128 GB board). On Linux the amdgpu GTT pool can be raised to roughly 108-120 GB with ttm.pages_limit / amdgpu.gttsize boot parameters. The Windows figure is used here as the conservative one. Measured GPU bandwidth is ~212-215 GB/s of the 256 GB/s theoretical, so estimates here run a little high.
- Power under load
- 133 W (third party measured) — Jeff Geerling's ai-benchmarks: Framework Desktop mainboard at 133 W running Llama 3.1 70B in Ollama (98 W on gpt-oss-120b via Vulkan). Meter method not stated.
- Availability
- System $3,449 (out of stock when checked); mainboard alone $3,149 on pre-order. Launched at $1,999 in 2025; LPDDR5x prices rose since.
- Sources
- source 1, source 2, source 3, source 4, source 5