Exact checkpoint example: Kimi-K2.5 is 595.178GB by official blob sizes; its 32B active parameters do not determine storage memory.
Large-model boundary: the MiniMax-M2.5 guide shows why a 230.134GB checkpoint makes 240GB unsafe and 256GB only a constrained file fit.
Short answer: your GPU memory label does not map to one maximum model size. The exact artifact, quantization, context length, KV-cache precision, runtime allocations and concurrency determine what fits. As a starting point, 4GB is mainly a small-model or offload tier; 8GB can support many 4B and selected 7B/8B configurations; 12–16GB opens more practical 7B, 8B and 14B paths; 24–32GB is the main single-GPU range for many quantized 32B workloads; and 48GB or more is where selected 70B-class full-GPU configurations become plausible. Verify the exact file before buying.
Calculate your actual setup: enter parameter count, precision, context and concurrency in the LLM VRAM calculator. Then compare the result with the exact artifact and your runtime’s own memory estimate.
GPU memory to model-size starting points
| Available GPU memory | Reasonable first models to investigate | Main constraint to verify |
|---|---|---|
| 4GB | Selected low-bit 4B; constrained 7B with a small artifact or offload | Cache and runtime allocations can consume the small remaining margin |
| 6GB | Many 4B paths; selected 7B/8B low-bit configurations | A file that nearly fills 6GB leaves little working room |
| 8GB | Broad 4B coverage; many 7B paths; selected 8B 4-bit paths | Some official 7B/8B Q8 artifacts already exceed 8GB before overhead |
| 12GB | Practical 7B/8B; selected 14B low-bit configurations | Context, cache precision and desktop use can turn a nominal fit into an out-of-memory failure |
| 16GB | Broad 7B/8B; many 14B paths; constrained 32B with aggressive quantization or offload | A typical 32B 4-bit artifact can exceed this pool |
| 24GB | Flexible 14B; many 24B, 27B and 32B 4-bit paths | Exact artifacts, higher quantizations and long context may require more room; 70B usually needs offload |
| 32GB | Broader 24B, 27B and 32B quantization/context choices | Common 70B 3-/4-bit artifacts can exceed 32GB before cache |
| 48GB | 32B with substantial headroom; selected 70B 4-bit paths | A 44GB Q4_K_M 72B example is tight after cache and runtime allocations |
| 64GB+ | More 70B 4-/5-bit options, longer context or concurrency | Shared memory, software support and runtime placement still matter |
Planning the middle tier? Use the 24B hardware guide or the 27B VRAM requirements guide for artifact-backed boundaries before choosing hardware.
Why parameter count is only the first input
Raw weight memory is approximately parameters × bits per weight ÷ 8. That gives 3.5GB for exactly 7B parameters at 4-bit precision and 35GB for exactly 70B parameters at 4-bit precision, but real artifacts can be larger because formats include metadata and mixed-precision tensors. Inference also needs KV cache, runtime buffers and safety headroom.
- Quantization: Q4, Q5, Q6 and Q8 are not interchangeable, and labels alone do not establish file size or quality.
- Context: the KV cache grows with token slots and can consume gigabytes even when the model weights fit.
- Concurrency: multiple active sequences can multiply cache and workspace needs.
- Runtime: implementations place weights, cache and buffers differently.
- Other memory users: a display, browser, desktop and adjacent AI pipeline may share the same pool.
Official artifact examples behind the table
| Model example | Official displayed artifact evidence | Detailed guide |
|---|---|---|
| Qwen2.5-3B / Qwen3-4B | 3B: 6.17GB displayed BF16/FP16 shards; 4B: 2.50GB Q4_K_M through 4.28GB Q8_0 | 3B/4B requirements |
| Qwen2.5-7B | 4.68GB Q4_K_M; 8.10GB Q8_0 | 7B requirements |
| Qwen3-8B | 5.03GB Q4_K_M; 8.71GB Q8_0 | 8B requirements |
| Meta Llama 2 13B | 26.03GB displayed FP16 shards; 3.125GiB illustrative FP16 cache at 4K | 13B requirements |
| Qwen2.5-14B | 5.77GB Q2_K through 15.7GB Q8_0 | 14B requirements |
| Devstral / Magistral 24B | 14.3GB Q4_K_M; 16.8GB Q5_K_M; 25.1GB Q8_0 official GGUF examples | 24B Mistral model map |
| Gemma 3 27B IT | 17.2GB official QAT Q4_0 model file; separate 858MB vision projector | Gemma 3 27B requirements |
| Qwen2.5-32B | 19.9GB Q4_K_M; 34.8GB Q8_0 | 32B requirements |
| Qwen2.5-72B | 44GB Q4_K_M; 77.5GB Q8_0 | 70B requirements |
Dedicated VRAM, unified memory and system RAM are different paths
A discrete GPU’s VRAM is a dedicated pool. Apple silicon and some integrated systems use memory shared with the operating system and applications, so the headline capacity is not all available to the model. CPU offload can place part of a workload in system RAM, but it changes the data path and can reduce performance. Two GPUs also remain separate pools unless the runtime explicitly partitions tensors between them.
Use the CPU-offload guide and multi-GPU guide before treating system RAM or two capacity labels as equivalent to one larger GPU.
A safer way to choose the model
- Choose the exact repository and artifact, not only a parameter class.
- Check the artifact’s displayed size and format.
- Set the context, cache precision and expected simultaneous sequences.
- Add runtime allocations plus space for the display and other applications.
- Confirm GPU architecture, operating system, driver and runtime support.
- Test one representative long request and evaluate output quality, not merely successful loading.
- Only then compare complete-system cost, power, noise and upgrade options.
Turn the fit estimate into a decision
Compare quantizations · Compare GPU and system paths · Run a repeatable workload test · Share a measured result
Track hardware that matches your model
Save a private browser-only watchlist and choose a focused RSS feed on the market-follow page. No account or email address is required.
Exact Qwen vision path: the Qwen2.5-VL 7B VRAM guide compares complete Q4/Q8 GGUF-plus-projector files with the official 16.584GB BF16 checkpoint.
Qwen vision family: use the Qwen2.5-VL 3B-to-72B VRAM guide for official checkpoint totals, architecture-specific cache geometry and image/video workload factors.
Vision-language family: the Qwen3-VL VRAM guide maps official 2B, 4B, 8B and 32B multimodal artifacts to practical memory tiers.
Compact vision case: the Qwen3-VL 4B VRAM guide maps official multimodal artifacts to 4GB, 6GB, 8GB and 12GB decisions.
Smallest Qwen3 vision case: the Qwen3-VL 2B VRAM guide maps official Q4, Q8 and BF16 artifacts to compact hardware.
MoE vision example: the Qwen3-VL 30B-A3B guide separates active routed parameters from stored weights and complete projector memory.
Flagship vision example: the Qwen3-VL 235B-A22B guide separates active routed parameters from exact 142.906GB Q4 and 471.340GB BF16 storage.
Qwen vision MoE branch: the Qwen3-VL MoE guide maps official 30B-A3B and 235B-A22B artifacts from 24GB through 640GB.
Sources and limits
- Official Qwen3-4B artifact and configuration sources
- Official Qwen2.5-7B artifact and configuration sources
- Official Qwen3-8B artifact and configuration sources
- Official Qwen2.5-14B artifact and configuration sources
- Official Devstral and Magistral 24B GGUF sources
- Official Gemma 3 27B and sourced 32B comparison evidence
- Official Qwen2.5-32B artifact and configuration sources
- Official Qwen2.5-72B artifact and configuration sources
- HighVRAM methodology
Reviewed October 25, 2026. The table is a planning map built from the linked official examples, not a compatibility promise for every model with the same parameter count. HighVRAM has not benchmarked every tier. Fit does not establish speed or output quality.
Choosing above the mainstream tier? Compare what 24GB, 32GB and 48GB GPUs can run before selecting a specific card.
New to local inference? Use the Can my PC run a local LLM? checklist before choosing a model or upgrade.
Planning a 70B-class workload? Compare 48GB, 64GB, 80GB and 96GB memory paths before choosing a workstation or server accelerator.
Running a mixture-of-experts model? Read why active parameters do not replace total-weight memory, with official Qwen3-30B-A3B and gpt-oss examples.
Choosing gpt-oss-20b or 120b? Read the official memory and hardware-support comparison.
Choosing Qwen3-30B-A3B? Read the official-artifact and cache guide before treating A3B as a 3B memory requirement.
Buying for Qwen3-30B-A3B or a dense 32B model? Compare 24GB, 32GB and 48GB hardware paths.
Exact-model example: Llama 3.1 70B
See the Llama 3.1 70B VRAM requirements guide for a worked example showing how precision and the model’s published 128K context can change the answer beyond a simple parameter-count table.
Exact-model example: Llama 3.3 70B
Use the Llama 3.3 70B VRAM guide to see why its newer model version does not create a smaller memory tier than 3.1.
Exact-model example: DeepSeek-R1-Distill-Llama-70B
Size the distilled 70B checkpoint without confusing it with DeepSeek’s much larger 671B mixture-of-experts model.
Exact-model example: DeepSeek-R1-Distill-Qwen-32B
Size the dense 32B distillation without confusing it with DeepSeek’s 70B distillation or 671B MoE model.
DeepSeek-R1 distill model map
Compare all six 1.5B-through-70B distill checkpoints using DeepSeek’s official repository totals and cautious compact-quant planning lanes.
Exact-model example: DeepSeek-R1-Distill-Qwen-14B
Size the distilled 14B model with official checkpoint evidence and 8K/32K/128K cache arithmetic.
Exact comparison: DeepSeek 7B vs 8B
Compare the Qwen and Llama distill variants using official repository totals and model-specific cache arithmetic.
DeepSeek 1.5B on small devices
The exact DeepSeek-R1-Distill-Qwen-1.5B guide covers 2GB–8GB paths and long-context cache limits.
Exact Qwen3-8B planning
The Qwen3-8B VRAM guide adds official GGUF sizes, cache arithmetic and 8GB/12GB/16GB decisions.
Exact Qwen3-14B planning
The Qwen3-14B VRAM guide adds official file sizes, cache arithmetic and 12GB/16GB/24GB decisions.
Exact Qwen3-32B planning
The Qwen3-32B VRAM guide adds official files, cache arithmetic and 24GB/32GB/48GB decisions.
Exact Qwen3-4B planning
The Qwen3-4B VRAM guide adds official files, cache arithmetic and 4GB/6GB/8GB decisions.
Dense Qwen3 size map
The Qwen3 family VRAM guide compares official 4B, 8B, 14B and 32B artifacts from 4GB through 48GB.
Planning for Gemma? Use the exact Gemma 3 27B VRAM requirements guide for official file sizes, vision overhead and capacity tiers.
Comparing one model family? The Gemma 3 VRAM guide maps official 1B, 4B, 12B and 27B QAT files from 2GB through 32GB+ memory tiers.
Exact 12B example: use the Gemma 3 12B VRAM guide for official model/projector sizes and 8GB, 12GB and 16GB planning.
Exact 4B example: use the Gemma 3 4B VRAM guide for official model/projector sizes and 4GB, 6GB and 8GB planning.
Exact low-memory example: use the Gemma 3 1B VRAM guide for official file size and 1GB, 2GB and 4GB planning.
Exact Microsoft example: use the Phi-4 14B VRAM guide for official quantization sizes and 12GB, 16GB and 24GB planning.
Exact 24B example: use the Magistral Small VRAM guide for official file sizes and Mistral’s chat-template, context and vision-path warnings.
New exact-model guide: Devstral Small 1.1 24B VRAM requirements maps Mistral’s official files to 16GB, 24GB, 32GB and 64GB+ paths.
Exact MoE guide: Mixtral 8x7B VRAM requirements combine official checkpoint/configuration evidence with labeled community GGUF sizes and cache arithmetic.
Exact multimodal guide: Llama 3.2 11B Vision VRAM requirements separate official weights, text cache, vision processing and community projector packaging.
Exact 90B multimodal guide: Llama 3.2 90B Vision VRAM requirements separate official weights, language cache, vision allocations and multi-device placement.
Llama 3.2 family map: compare official 1B, 3B, 11B Vision and 90B Vision checkpoint sizes and cache costs before choosing a tier.
Exact Mistral guide: Mistral Small 3.2 24B VRAM requirements combine official multimodal checkpoint evidence with labeled community quantizations.
Mistral 24B model map: the Small 3.2 vs Devstral vs Magistral guide routes equal-sized artifacts into the correct general, coding or reasoning workload.
Multimodal path: use the vision LLM VRAM comparison for Gemma 3, Llama 3.2 Vision and Mistral Small 3.2 projector/encoder memory.
Current hybrid vision case: Qwen3.5-27B maps first-party BF16 and labeled community GGUF evidence to 24GB–80GB decisions.
Current sparse multimodal case: Qwen3.5-35B-A3B compares official GPTQ/BF16 artifacts and labeled community GGUF files without confusing 3B active with stored weights.
Current 9B vision case: Qwen3.5-9B maps first-party BF16 and labeled community Q4/Q8-plus-projector evidence to 8GB–32GB decisions.
Current compact vision case: Qwen3.5-4B maps first-party BF16 and labeled community Q4/Q8-plus-projector evidence to 4GB–16GB decisions.
Current small vision case: Qwen3.5-2B maps first-party BF16 and labeled community Q4/Q8-plus-projector evidence to 2GB–8GB decisions.
Current sub-1B vision case: Qwen3.5-0.8B maps first-party BF16 and labeled community Q4/Q8-plus-projector evidence to 1GB–4GB decisions.
Qwen3.5 model map: compare 0.8B, 2B, 4B, 9B, 27B and 35B-A3B across 1GB–96GB planning tiers without confusing active parameters or file size with peak memory.
Current 122B MoE case: Qwen3.5-122B-A10B maps exact official checkpoints and labeled community files to 80GB–320GB multi-device decisions.
397B sparse model: the Qwen3.5-397B-A17B VRAM guide covers exact 4-bit, FP8/Q8 and BF16 file paths plus multi-device headroom.
Checking Llama 4 specifically? Use the Scout and Maverick VRAM guide instead of treating their 17B active label like a dense 17B model.
Current MoE example: the GLM-4.7 VRAM guide shows why Flash BF16 starts at a 62.444GB checkpoint while the full FP8 and BF16 repositories require 362.064GB and 716.681GB before runtime overhead.
Flagship MoE example: the DeepSeek-V3.2 guide maps its exact 689.483GB official checkpoint across 640GB, 768GB and larger distributed tiers.