BuildsGPUsModelsGuidesMethodologyNewsletter

What local AI models can the Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB run?

192 GB LPDDR5X-8533, 273 GB/s, W TDP. Specs: TechPowerUp (accessed 2026-10-15). Price: low $6,799 (new) source, 2026-10-16

Model (Q4)FileFits?Speed estimate
Qwen3-0.6B0.48 GB✅~432 tok/s
Llama-3.2-3B-Instruct2.02 GB✅~103 tok/s
gemma-3-4b-it2.49 GB✅~83 tok/s
Qwen3-4B2.5 GB✅~83 tok/s
Mistral-7B-Instruct-v0.34.37 GB✅~47 tok/s
Qwen2.5-Coder-7B-Instruct4.68 GB✅~44 tok/s
Llama-3.1-8B-Instruct4.92 GB✅~42 tok/s
Qwen3-8B5.03 GB✅~41 tok/s
NVIDIA-Nemotron-Nano-9B-v26.53 GB✅~32 tok/s
gemma-3-12b-it7.3 GB✅~28 tok/s
Mistral-Nemo-Instruct-24077.48 GB✅~28 tok/s
Qwen2.5-Coder-14B-Instruct8.99 GB✅~23 tok/s
DeepSeek-R1-Distill-Qwen-14B8.99 GB✅~23 tok/s
Qwen3-14B9 GB✅~23 tok/s
phi-49.05 GB✅~23 tok/s
gpt-oss-20b11.67 GB✅~103 tok/s
Mistral-Small-3.2-24B-Instruct-250614.33 GB✅~14 tok/s
Devstral-Small-250514.33 GB✅~14 tok/s
gemma-3-27b-it16.55 GB✅~13 tok/s
Qwen3-Coder-30B-A3B-Instruct18.56 GB✅~103 tok/s
Qwen3-30B-A3B18.63 GB✅~103 tok/s
Qwen3-32B19.76 GB✅~10 tok/s
Qwen2.5-Coder-32B-Instruct19.85 GB✅~10 tok/s
DeepSeek-R1-Distill-Qwen-32B19.85 GB✅~10 tok/s
Llama-3.3-70B-Instruct42.52 GB✅~5 tok/s
DeepSeek-R1-Distill-Llama-70B42.52 GB✅~5 tok/s
Qwen2.5-72B-Instruct47.42 GB✅~4 tok/s
gpt-oss-120b62.84 GB✅~76 tok/s
GLM-4.5-Air73.5 GB✅~26 tok/s
Qwen3-235B-A22B142.15 GB✅~16 tok/s

Speed calculator estimate

Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works

Related