Which GPU fits the work you actually do?

Small dense open-weight reference. The publisher lists 4.0B parameters. This site's Q4_K_M planning estimate is about 2.4GB of weights before its allowances; this is not a measured file size. Check the quantized pack, available memory and runtime, including on integrated graphics.

Earlier 8B text checkpoint for capacity comparisons. Meta's 2024 8B instruct checkpoint uses the Llama 3.1 Community License. Llama 3.2 has 1B/3B text and 11B/90B vision variants, rather than an 8B text successor. The site's Q4_K_M weight estimate is about 4.9GB before its allowances; an estimated 8GB fit still depends on context, runtime and the actual file.

Dense multimodal 27B reference. Dense 27B with a vision encoder. This site's Q4_K_M estimate is about 16.4GB of weights before its allowances; it exceeds 16GB and suggests checking a 24GB configuration. Actual vision/cache/runtime allocations and quantized files vary. The native checkpoint publication day is unverified.

Dense multimodal open-weight reference. Google's card lists 30.7B total parameters despite the 31B name. This site's Q4_K_M weight estimate is about 18.6GB; Q4_0 and specific downloadable files need separate sizing. The house heuristic marks 24GB tight and 32GB comfortable, which does not guarantee a context length or runtime fit.

Dense model for 12–16GB planning comparisons. The publisher lists 14.8B parameters despite the 14B name. The site's Q4_K_M weight estimate is about 9.0GB and its required-memory estimate about 10.9GB. Its heuristic labels 12GB tight and 16GB comfortable; actual file, cache and runtime memory can change the result.

27B open-weight fit example. Earlier Gemma open-weight reference under Gemma Terms of Use. Using 27B as the marketed parameter count, the site estimates about 16.4GB of Q4_K_M weights before its allowances, already above 16GB. A 24GB configuration is a planning candidate, not a measured fit or a verified downloadable file size.

Mixture-of-experts capacity reference. The publisher lists 30.5B total and 3.3B active parameters. Active count is not the weight-storage requirement: the site's Q4_K_M estimate uses the total and is about 18.5GB before its allowances. Full residency, CPU offload and expert routing depend on the backend; active count alone proves no speed advantage.

Dense model for 24–32GB planning comparisons. The publisher lists 32.8B parameters despite the 32B name. The site's Q4_K_M estimate is about 19.9GB of weights and 22.9GB with its allowances: a 24GB planning candidate, not 16GB. The house heuristic labels 24GB tight and 32GB comfortable; actual context, file and runtime allocations remain unmeasured.

Open multimodal agentic distill for 24–32GB class cards. The publisher lists about 29.6B total including the vision encoder. Official GGUF documentation lists roughly 17GB and 20GB language packs, a separate 1.4GB perception pack for images and an optional 1.6GB drafter. Publisher examples target 24–32GB configurations; use the exact pack and runtime requirements rather than treating the generic Q4 estimate as a measured multimodal fit.

Larger text checkpoint for workstation capacity comparisons. The site's Q4_K_M estimate is about 42.4GB of weights and 47.7GB with its house allowances. Its heuristic labels a 48GB card or a 64GB unified-memory machine tight. The 75% unified-memory fraction and fixed overhead are planning assumptions; actual model loading, context and performance are untested.

Earlier large MoE for workstation capacity comparisons. The publisher lists 141B total parameters; the marketed active count is about 39B. For a fully resident model, storage includes all weights. The site's Q4 estimate is about 85.5GB, or 95.0GB with its house allowances: a tight fit on a 96GB workstation card under that heuristic. Offload, context, runtime and multiple devices can change the result; no speed or stability guarantee is established.

Large open-weight reasoning example. The publisher lists 235B total and 22B active parameters. The site's Q4 weight estimate is about 142.5GB before its allowances and assumes storage for the full model. Active count alone proves neither memory fit nor speed; runtime/offload and actual quantized files require separate checks.

Large open-weight coding example. The publisher lists 480B total and 35B active parameters. The site's Q4 weight estimate is about 291GB and its required-memory estimate about 321GB. A 512GB unified-memory configuration is tight under its 75%/80% thresholds, not a verified deployment; Apple's announced 512GB M5 Ultra Studio configuration is due in late October 2026. Backend support, offload, actual files and context matter.

Earlier open-weight DeepSeek reference. The native publisher card lists 284B total / 13B active and mixed FP4/FP8 weights under MIT. The site's generic Q4 estimate is about 172GB before its allowances, not the size of that native mixed-precision pack. This is the earlier text-only V4 checkpoint; hosted V4 aliases now route to V4.1 and do not identify these weights. V4.1's different backbone and conditional-memory architecture must not inherit this sizing.