Framework Laptop 16 DIY Edition (2025) with RTX 5070 12GB: which AI models can it run?
Estimated from the specifications below. Choose the memory size you have or plan to buy, then read the table for that size.
Specifications
- Brand
- Framework
- Year
- 2025
- Chip
- AMD Ryzen AI 5 340, Ryzen AI 7 350 or Ryzen AI 9 HX 370
- Processor
- Ryzen AI 5 340, AI 7 350 or AI 9 HX 370
- Graphics
- NVIDIA GeForce RTX 5070 Laptop GPU (12GB)
- Memory type
- DDR5-5600 SO-DIMM
- Memory options
- 96 GB
- Graphics memory
- 12 GB
- Memory bandwidth
- 384 GB/s
Sources: https://frame.work/laptop16?tab=specs
Framework states 12GB GDDR7, 128-bit bus and 384GB/s for this graphics module. Memory is two SO-DIMM sockets with "up to 96GB"; smaller sizes are not itemised.
With 96 GB of memory
About 12 GB of it can hold a model.
| Model | Fit | Needs (GB) | Writing speed (tokens per second) | Speed | Quality | Basis |
|---|---|---|---|---|---|---|
| Llama 3.2 3B llama3.2:3b |
Fits | 5 | 67 to 125 | Fast | not measured yet | Estimated |
| Qwen3 4B qwen3:4b · thinks first |
Fits | 6 | 54 to 100 | Fast | not measured yet | Estimated |
| Gemma 3 4B gemma3:4b |
Fits | 6.7 | 41 to 76 | Fast | not measured yet | Estimated |
| Mistral 7B v0.3 mistral:7b |
Fits | 7.8 | 31 to 57 | Fast | not measured yet | Estimated |
| Llama 3.1 8B llama3.1:8b |
Fits | 8.3 | 27 to 51 | Fast | not measured yet | Estimated |
| Qwen3 8B qwen3:8b · thinks first |
Fits | 8.9 | 26 to 48 | Fast | not measured yet | Estimated |
| Gemma 3 12B gemma3:12b Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 15.9 | 2.5 to 4.6 | Slow | not measured yet | Estimated |
| DeepSeek-R1 Distill Qwen 14B deepseek-r1:14b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 13.7 | 2.2 to 4.2 | Slow | not measured yet | Estimated |
| Phi-4 14B phi4:14b Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 13.9 | 2.2 to 4.1 | Slow | not measured yet | Estimated |
| Qwen3 14B qwen3:14b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 13.4 | 2.2 to 4 | Slow | not measured yet | Estimated |
| gpt-oss 20B gpt-oss:20b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 16.5 | 5 to 9.3 | Moderate | not measured yet | Estimated |
| Gemma 3 27B gemma3:27b Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 27.2 | 1.2 to 2.2 | Slow | not measured yet | Estimated |
| Qwen3 30B (A3B) qwen3:30b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 22.6 | 5.9 to 11 | Unclear could be Moderate or Fast | 9.3/10 | Estimated |
| Qwen3 32B qwen3:32b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 26.3 | 1 to 1.9 | Slow | not measured yet | Estimated |
| Llama 3.1 70B llama3.1:70b Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 51.5 | 0.5 to 0.9 | Slow | not measured yet | Estimated |
| gpt-oss 120B gpt-oss:120b · thinks first Part of the model runs from system memory, which is much slower. |
Partial offload (slow) | 70.5 | 4.3 to 7.9 | Unclear could be Slow or Moderate | not measured yet | Estimated |
Models marked "thinks first" can write out their reasoning before they answer, so a job takes longer than the speed alone suggests.
Speed is the speed of writing an answer, shown as a range. Reading speed is not estimated. How the estimates work · How close they are · Back to the finder