FitMyModel

What local AI models run on 48 GB of VRAM?

Last updated

We earn from qualifying purchases via affiliate links. Disclosure.

10 of 13 ranked open models fit in 48 GB: 10 with 4 GB or more to spare and 0 tight. The best ranked is Gemma 4 31B (19–20 GB).

ModelLMArenaNeedsFitPages
Gemma 4 31BGoogle · Q4_K_M
 #76 19–20 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 24–32 GB GPU tier
Gemma 4 26B-A4B (MoE)Google · default tag
 #100 16 GB Fits Ryzen AI Max+ 395 (64 GB) · 24–32 GB GPU tier
Qwen3.8 27BAlibaba Qwen · Q4_K_M
 #102 18 GB Fits Ryzen AI Max+ 395 (64 GB) · 24–32 GB GPU tier
Qwen3.5 27BAlibaba Qwen · Q4_K_M
 #155 17–20 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 24–32 GB GPU tier
Qwen3.5 35B-A3B (MoE)Alibaba Qwen · Q4_K_M
 #172 22 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 24–32 GB GPU tier
Mistral Small 3.2 24BMistral AI · Q4_K_M
 #212 15 GB Fits Ryzen AI Max+ 395 (64 GB) · 16 GB GPU tier
Llama 3.3 70BMeta · Q4_K_M
 #266 43 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · Unified 64 GB tier
gpt-oss 20BOpenAI · MXFP4 (as shipped)
 #269 14 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 16 GB GPU tier
Phi-4 14BMicrosoft · Q4_K_M
 #322 9.1 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 8–12 GB GPU tier
Llama 3.1 8BMeta · Q4_K_M
 #345 4.9 GB Fits Ryzen AI Max+ 395 (64 GB) · Guide · 8–12 GB GPU tier

Fit compares the Ollama library download size with 48 GB. "Fits" means 4 GB or more spare and "Tight" means under 4 GB; both bands are our estimate, and tight models need a short context. Ranks from LMArena Text, checked Oct 27, 2026, 2:16 PM CT. No speeds are listed because we have no sourced numbers.

Hardware with 48 GB for the GPU

Prices: Newegg checked Oct 27, 2026, about 12:35 PM CT; Minisforum checked Oct 27, 2026, 12:40 PM CT. New condition only.

Other memory sizes

← 32 GB 8 GB 12 GB 16 GB 24 GB 32 GB 48 GB 64 GB 96 GB 128 GB 64 GB →

Try the What fits my memory tool or browse every pairing on Can it run?

FAQ

What local AI models run on 48 GB of VRAM?

10 of the 13 leaderboard-ranked open models we track fit in 48 GB: Gemma 4 31B (19–20 GB), Gemma 4 26B-A4B (MoE) (16 GB), Qwen3.8 27B (18 GB), Qwen3.5 27B (17–20 GB), Qwen3.5 35B-A3B (MoE) (22 GB), Mistral Small 3.2 24B (15 GB), Llama 3.3 70B (43 GB), gpt-oss 20B (14 GB), Phi-4 14B (9.1 GB), Llama 3.1 8B (4.9 GB). Sizes are Ollama library downloads; all leave 4 GB or more spare (our estimate).

What is the best-ranked local model for 48 GB?

Gemma 4 31B, #76 overall on LMArena Text (checked Oct 27, 2026, 2:16 PM CT), needs 19–20 GB.

What is the largest model that fits in 48 GB?

Llama 3.3 70B at 43 GB (Q4_K_M). The runtime and context need roughly 1–4 GB more, which is our estimate.

Does 48 GB of unified memory count as 48 GB of VRAM?

No. On a Ryzen AI Max+ 395 system AMD says the GPU can use up to 48 GB of a 64 GB system and up to 96 GB of a 128 GB system, so use the GPU share when you read this page.