Can your GPU or Mac run this AI model?
New open models ship every week. Pick your graphics card or Mac and RAM, and each model's own checker answers from the developers' published requirements and benchmarks.
Short answer: 5 model sizes on GeForce RTX 3060 (12 GB) + 32 GB RAM; 7 model sizes on GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM. At the file choices below, every model layer fits in GPU memory. These are RigCheck capacity estimates for llama.cpp at a 32K context with no image input, not measured runs.
Required file choices: With GeForce RTX 3060 (12 GB) + 32 GB RAM, Gemma 4 12B (Q5_K_M), Gemma 4 E2B (BF16), Gemma 4 E4B (Q8_0), LFM2.5-24B-A2B (UD-IQ3_XXS), Qwen3.8-27B (UD-IQ2_XXS) fit with every layer in GPU memory; with GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM, Gemma 4 12B (Q8_0), Gemma 4 26B-A4B (UD-Q4_K_XL), Gemma 4 31B (IQ4_XS), Gemma 4 E2B (BF16), Gemma 4 E4B (BF16), LFM2.5-24B-A2B (Q6_K), Qwen3.8-27B (UD-Q4_K_XL) fit with every layer in GPU memory. Input embeddings and host buffers can still use system RAM. Larger models need system RAM or a Mac; image and video models use planning budgets and published recipes. Every model: the model list.
What can my hardware run?
Open in a full page ↗Loading private checker…
14 models · 1 engine · 5 VRAM pages · Latest: Recursive Flint, added
Start from your hardware
By VRAM size
| Graphics card | VRAM | Language models | Image models | Video (published recipes) |
|---|---|---|---|---|
| RTX 3060 12 GB | 12 GB | 5 with every layer on the GPU · 7 more with system RAM, caveats or an experimental path · of 14 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): Above the all-components-resident planning budget; offload peak unverified | Hardware memory matches a published video recipe; exact clip peak unverified |
| RTX 4070 Ti | 12 GB | 5 with every layer on the GPU · 6 more with system RAM, caveats or an experimental path · of 13 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): not covered | not covered |
| RTX 4060 Ti 16 GB | 16 GB | 7 with every layer on the GPU · 5 more with system RAM, caveats or an experimental path · of 14 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): not covered | Hardware memory matches a published video recipe; exact clip peak unverified |
| RTX 3090 | 24 GB | 7 with every layer on the GPU · 5 more with system RAM, caveats or an experimental path · of 14 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): Above the all-components-resident planning budget; offload peak unverified | Hardware memory matches a published video recipe; exact clip peak unverified |
| RTX 4090 | 24 GB | 7 with every layer on the GPU · 5 more with system RAM, caveats or an experimental path · of 14 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): Above the all-components-resident planning budget; offload peak unverified | Hardware memory matches a published video recipe; exact clip peak unverified |
| RTX 5090 | 32 GB | 7 with every layer on the GPU · 5 more with system RAM, caveats or an experimental path · of 14 sizes | FLUX 3 Image Klein: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; Qwen-Image-2.1 (Base and Turbo): Within the all-components-resident planning budget; allocation unverified | Hardware memory matches a published video recipe; exact clip peak unverified |
With 64 GB of system RAM, in the finder's quick mode. Language models are counted in sizes (Gemma 4 counts five); sizes the checker does not cover and models without released weights are left out. Image and video: Qwen-Image (a planning budget): base model, INT8, 512×512, batch one, no reference images; Kolibri-1: the official FP8 route at a 32K context; MiniMax H3: local H3-Base FL2VA recipes, assuming an NVMe drive; Nemotron 3.5: Bolt and Super VL on one card with the RAM given; FLUX 3 Image Klein: Black Forest Labs' int8 ComfyUI route on the card. Language-model answers are RigCheck capacity estimates (Strata: its published requirements; Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash: published routes and Unsloth memory tables), not measured setups. All graphics cards · Macs.
Fits wholly on one gaming GPU
Sizes that sit entirely in the memory of a card with 24 GB or less, at the cheapest card in our sample; sizes that need system RAM are left out.
- Gemma 4 12B: Q5_K_M wholly in GPU memory with GeForce RTX 3060 (12 GB) + 32 GB RAM (capacity estimate)
- Gemma 4 E2B: BF16 wholly in GPU memory with GeForce RTX 3060 (12 GB) + 32 GB RAM (capacity estimate)
- Gemma 4 E4B: Q8_0 wholly in GPU memory with GeForce RTX 3060 (12 GB) + 32 GB RAM (capacity estimate)
- LFM2.5-24B-A2B: UD-IQ3_XXS wholly in GPU memory with GeForce RTX 3060 (12 GB) + 32 GB RAM (capacity estimate)
Newly added
- Recursive Flint: added Oct 23, 2026. Will Recursive's 40B-A2.5B coding model fit your graphics card, RAM or Mac with llama.cpp or Ollama? Real GGUF file sizes, the KV cache for your context and the file to download.
- LFM2.5-24B-A2B: added Oct 22, 2026. Will Liquid AI's 24B mixture-of-experts model fit your graphics card, RAM or Mac with llama.cpp or Ollama? Real GGUF file sizes, the KV cache for your context and the file to download.
- DeepSeek V4.1 Flash: added Oct 18, 2026. Match your Mac, Spark or PC to the routes DeepSeek V4.1 Flash's sources name: DwarfStar on one or two 128 GB machines (1017 or September Q2), Strata's 1017 sets on a 12 GB RTX card with 46 GB of RAM, Unsloth's experimental 1017 GGUF, and the real file sizes.
Image and video
Judged by planning budgets and published recipes, not by a measured peak.
- FLUX 3 Image Klein: GeForce RTX 3060 (12 GB) + 32 GB RAM: Matches BFL's 8 GB guidance: the int8 files in ComfyUI (published route match). BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results.
- MiniMax H3: GeForce RTX 3060 (12 GB) + 32 GB RAM: Hardware memory matches a published video recipe; exact clip peak unverified (published recipe match). Matching recipes: SGLang Recipe A (lab): one card, DiT, text encoder and VAE streamed from the checkpoint file on an NVMe drive (Recipe A), BF16; measured at 864×480 (at the native 768P the cookbook says to use video_vae=24 if the decode runs out of memory, not measured); lab run with VRAM capped at 12 GiB and a simulated 32 GB host: shows the memory fits, not your card's speed; a real 32 GB machine re-reads the weights from disk on every step. H3-Base FL2VA, source recipe canvases only. Disk-streaming routes assume an NVMe drive. Host-RAM and lab assumptions remain in the full checker; full 2K needs an API.
- Qwen-Image-2.1: GeForce RTX 3060 (12 GB) + 32 GB RAM: Above the all-components-resident planning budget; offload peak unverified (planning budget). Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction.
Recently updated
- Qwen3.8-27B: Unsloth re-uploaded four of the checker's files on November 11: UD-Q5_K_XL now 21.0 GB, UD-Q4_K_XL 17.7 GB, UD-IQ4_XS 14.4 GB, UD-Q2_K_XL 9.9 GB (each 0.1 GB larger)
- Qwen3.8-27B: UD-IQ2_XXS: Unsloth re-uploaded it on November 10; now 7.2 GB (was 7.3 GB), about 11.5 GB at a 32K context
- Kolibri-1: llama.cpp: its own source now includes the kolibri1 architecture; the GGUF recipe still uses the converter's patched build, and loading that file in unpatched llama.cpp is not verified, so the checker keeps the patched route
- MiniMax H3: File sizes: dropped the ComfyUI 0.38.0 note for the 16.0 GB pruned W6A8 transformer, because the ComfyUI guide no longer mentions that build (the file is still in Comfy-Org's list)
- Kolibri-1: Official route update: Aleph Alpha's model card now says vLLM supports Kolibri natively from version 0.33.0, with its built-in hf reasoning and tool parsers; the checker's official command uses that, and the Aleph Alpha plugin is only for older vLLM. Hardware, files and numbers are unchanged
How we check
- Numbers come from the model's own repository, model card or announcement, and every one links to its source.
- Official measurements, official estimates, community reports and RigCheck's own estimates are labeled differently.
- Each page shows when it last changed and when its numbers were last checked against the sources.
RigCheck is independent and not affiliated with any model developer. Found a wrong number? Email hello@localrigcheck.com with the source. How RigCheck works.
FAQ
How do I know if my GPU can run a model?
Pick your card and RAM in the finder above, or open your graphics card's page. Each answer comes from that model's own checker, and says which kind it is: the llama.cpp checkers add up the file, the context cache and buffers against your memory; the Strata checker applies Strata's published requirements; Kolibri-1's matches official minimum configurations and a community recipe; image models use a planning budget and video models match published recipes. The full checker on each model page takes your exact setup.
What does "with caveats" mean?
The model fits only with part of it in system RAM (slower), or the memory peak is not measured, or the route is one the checker marks as conditional. The checker says which; RigCheck never shows a caveat as a plain yes.
Can a Mac run these models?
Many of them: on a Mac the graphics and the processor share one pool of unified memory, so the finder judges a Mac by its total memory. Open the page for your Mac's memory size or pick your Mac in the finder; the Strata engine runs on Apple Silicon only experimentally.
Are these numbers measured?
Mostly not. The llama.cpp checkers' memory answers are RigCheck capacity estimates from the published file sizes; Strata's follow its published requirements, Kolibri-1's match official minimum configurations or a community recipe, and image and video answers are planning budgets or published recipes. Speeds are labeled as an official measurement, an official estimate or a community report; a measurement names the card and setup it was run on. How RigCheck works.
Which models does RigCheck cover?
Newly released open-weights models and a few popular mid-size ones, each with its own checker. The model list shows them all with their sizes, weights status, runtimes and license.