Best local LLMs for 32 GB RAM or VRAM

32 GB: weight budgets

Compared with 24 GB, this tier adds 29 weight-budget candidates: gemma-4-26B-A4B-it, Cloudflare Clef 27B, Qwen3.8-27B EfficientThink SFT-SimPO, d3 27B, Swift 1.5 Qwen3.8-27B, ThinkingCap Qwen3.8-27B, Qwen3.8-27B Uncensored (JonathanColetti), Holo4-27B and more in the tables below. A smaller model can still be the better trade-off.

MMLU-Pro: models within the RAM budget

Knowledge and reasoning across many subjects. Published accuracy; different evaluation settings can change the result.

LiveCodeBench v6: models within the RAM budget

Solving coding problems. Version 6 only; published pass rates can vary with generation settings.

All 124 model files within this tier’s RAM budget

Some files have scores only on other benchmarks. This complete size-ordered list keeps those candidates visible without assigning missing scores a zero.

Additional candidates with 32 GB dedicated VRAM

The 25.6 GB VRAM weight budget admits these additional files. They do not pass this page’s more cautious system-memory budget.

Computers near this memory tier

These are memory candidates, not tested model/runtime matches. Dedicated VRAM is used for GPU-fit planning; system RAM is not treated as VRAM.

• Beelink SER9 Max · Ryzen 7 260 · 32 GB system RAM
• GEEKOM A9 Max · 32 GB system RAM
• GEEKOM IT15 · 32 GB system RAM
• HP ProDesk 4 G1i · 32 GB system RAM
• Beelink SER10 Max · 32 GB system RAM
• GMKtec K15 · 32GB · 32 GB system RAM
• VZMORE AX9 Max · 32 GB system RAM
• BOSGAME AI 9 · 32 GB system RAM
• TOPCtec TA1 · 32 GB system RAM
• Apple Mac Studio M5 Max · 36GB · 36 GB unified memory
• Apple MacBook Pro 14 M5 Max · 36GB · 36 GB unified memory

How recommendations are calculated

We compare one benchmark at a time. A Pareto choice has no other scored model with both an equal or smaller weight file and an equal or higher score, with at least one strict improvement. The first choice is the frontier point nearest the chart’s top-left corner, using the same rounded memory scale and 0–100 score scale as the main chart. Other frontier choices follow, then other scored models by score.

Score per GB is shown for context; it does not choose the champion. Changing the budget or dataset can change the result. Scores are published point estimates, often for the source checkpoint rather than the selected quantization. They do not establish statistically significant differences or equal evaluation conditions.

RAM, VRAM and unified memory

Use one memory pool. System RAM is used for CPU inference; a dedicated GPU has separate VRAM. Do not add RAM and VRAM as if they were one GPU pool. Partial CPU offload needs a supported backend and can change speed. Unified memory is shared by CPU, GPU and the operating system; software can limit how much the GPU can allocate.

Our initial weight budget reserves 30% of system RAM or unified memory, with a minimum of 2 GB. For VRAM it reserves 20%, with a minimum of 1 GB. These are adjustable planning assumptions, not measured requirements. Context cache, runtime buffers, other programs, multimodal files and concurrent users can need more. A file that passes the budget check is a candidate, not a confirmed hardware-compatible configuration.

Check the model’s runtime and quantization support before downloading. See the Transformers cache documentation and llama.cpp backend documentation. GB uses decimal bytes; operating systems can report GiB instead.

Related memory tiers and tasks

All model records · Memory planning guide