llamaperf

LLM VRAM calculator

Pick your GPU or Mac and what you use models for. Every open-weight model the site knows is sized for your hardware, graded on fit, given a speed estimate from its memory bandwidth, and ranked by how well it matches the job. A Mac is graded against the memory macOS lets its GPU use, not all of it. Where community benchmarks exist on the same card, the measured number replaces the estimate.

On the M5 Ultra 512GB: Top match: GLM-5.3 753B · 40B active at Q3_K_S, Good fit. About 18.6 t/s (estimated). As of 22 Nov 2026.

Hardware

M5 Ultra·512 GB unified memory·1,229 GB/s

Balanced: quality first, then speed.

macOS lets the GPU use 384 of the 512 GB until you raise it.

Get a weekly email of new M5 Ultra 512GB reports and newly released models that fit it.

Email me new reports
0 perfect·3 good·0 marginal·0 too tight
Sort

People have run GLM-5.3 753B · 40B active on the M5 Ultra 512GB and reported their speeds.

Memory assumes an F16 KV cache at 32k context; Perfect is under 60% of the pool, Good under 85%, Marginal under 98%. Offload paths cap at Good. Speed is the memory bandwidth divided by the bytes read per token, at 65% efficiency (measured on this chip); a community median at the same quant replaces it. Click a row for the breakdown.

Did this help you choose a setup?

Optional feedback. No account needed.