What can a 32 GB graphics card run locally?
- Page updated
- Answers from each model's own checker
- Card by card, no formula
- Card memory from the vendors
What can a 32 GB graphics card run locally? Every answer below is the one RigCheck's own checker for that model gives for each 32 GB card in the finder, with 64 GB of system RAM. Nothing on this page is judged separately, and no answer is a formula from the memory size alone.
With 64 GB of system RAM, across the 3 graphics cards with 32 GB of VRAM in RigCheck's finder, the checkers have an answer for 14 of the 16 models, sizes and routes they cover on at least one of them. Placed wholly or partly on the card, by RigCheck's llama.cpp capacity estimates: Qwen3.8-Flash-Next in llama.cpp (1 of 3), Mistral Medium 3.5 in llama.cpp (1 of 3), Qwen3.8-27B in llama.cpp (2 of 3, without caveats or conditions on 1), Gemma 4 31B (2 of 3, without caveats or conditions on 1), Gemma 4 26B-A4B (2 of 3, without caveats or conditions on 1), Gemma 4 12B (2 of 3, without caveats or conditions on 1), Gemma 4 E4B (2 of 3, without caveats or conditions on 1), Gemma 4 E2B (2 of 3, without caveats or conditions on 1). On the processor in system RAM alone, the card unused: Qwen3.8-27B in llama.cpp (1 of 3), Gemma 4 31B (1 of 3), Gemma 4 26B-A4B (1 of 3), Gemma 4 12B (1 of 3), Gemma 4 E4B (1 of 3), Gemma 4 E2B (1 of 3). Answers that name no placement (Strata's published requirements, published route and recipe matches, planning budgets; some rest on system RAM alone, as the table says): Qwen3.8-Flash-Next in the Strata engine (all 3), Qwen-Image-2.1 (image generation) (1 of 3), MiniMax H3 (video + audio) (1 of 3), Nemotron 3.5 Bolt 74B-A9B (2 of 3), Nemotron 3.5 Super VL 120B-A12B (2 of 3), FLUX 3 Image Klein (int8, ComfyUI) (1 of 3). Which cards get an answer, and which answer, depends on the card and not only on its memory: the table lists them card by card. The llama.cpp language-model answers are capacity estimates and the Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash answers are published route matches, not measured setups.
32 GB of VRAM: model by model
Language models
| Model (checker) | With 64 GB RAM, card by card |
|---|---|
| Qwen3.8-Flash-Next in the Strata engine | ✓ Runs IQ3_S Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. On 2 of 3 cards: GeForce RTX 5090, Radeon AI PRO R9700 ! Experimental path IQ3_S Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. Pascal / Volta cards use Strata's separate CUDA 12 engine (the cuda12 app on Windows or Linux, or a Linux build), covered by the community "Older GPUs" guide. On 1 of 3 cards: Tesla V100 |
| Qwen3.8-Flash-Next in llama.cpp | ! Fits with caveats (capacity estimate) UD-Q2_K_XL · GPU and system RAM (53.1 GB in RAM) Flash-Next uses combined memory totals; per-card layer placement is not modeled. Close fits are uncertain until runtime allocation is checked. RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090 ✕ Does not fit Even the smallest file (UD-IQ1_S, 72.5 GB) needs about 76.1 GB in system RAM. This checker does not cover your graphics card, so it is not counted; its memory could change the answer. On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 |
| Mistral Medium 3.5 in llama.cpp | ! Fits with caveats (capacity estimate) IQ4_XS · GPU and system RAM (28 of 88 layers on the GPU): a dense model runs slowly from RAM RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090 – Not covered On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 |
| Qwen3.8-27B in llama.cpp | ✓ Fits (capacity estimate) UD-Q6_K_XL · GPU memory (input embeddings in system RAM) RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090 ✓ Fits (capacity estimate) UD-Q6_K_XL · GPU memory (input embeddings in system RAM) On this Radeon card this is a capacity estimate only: we found no source for this model with llama.cpp's own AMD builds. Ollama lists this card for its ROCm support, but its default package (ollama run qwen3.8:27b, 18 GB) is a different file from the one checked here, so compare its size with your memory first. RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Radeon AI PRO R9700 ! Fits with caveats (capacity estimate) Q8_0 · system RAM, CPU only RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100 |
| Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp | ✓ Best size: fits (capacity estimate) Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)) RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090 ✓ Best size: fits (capacity estimate) Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)) On this Radeon card these are capacity estimates only: we found no source for Gemma 4 with llama.cpp's own AMD builds. Ollama lists this card for its ROCm support, but its default packages (gemma4:31b 19–20 GB, gemma4:26b 16–19 GB, gemma4:12b 7.7–8.0 GB, gemma4:e4b 6.6–9.5 GB, gemma4:e2b 4.6–7.5 GB) are different files from the ones checked here. RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Radeon AI PRO R9700 ! Best size: fits with caveats (capacity estimate) Gemma 4 31B: fits with caveats (capacity estimate) (Q8_0, system RAM, CPU only); Gemma 4 26B-A4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 12B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E2B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only) RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100 |
| Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B) | ! Best size: a published route or recipe matches, with caveats Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 96 GB budget: UD-Q3_K_XL (65.8 GB) or NVFP4 (79 GB)) Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. Longer contexts and video need more. On 1 of 3 cards: GeForce RTX 5090 ! Best size: a published route or recipe matches, with caveats Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches Unsloth's short-context budget: about 32 GB of RAM or unified memory); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 64 GB budget: UD-IQ1_M (46.1 GB) or UD-IQ2_XXS (51.1 GB)) Nemotron 3.5 Bolt 74B-A9B: We found no source covering AMD GPUs for this model, so only system RAM is counted here. Unsloth budgets around 32 GB of RAM or unified memory for short-context use (48–64 GB for more headroom) with its 22 GB IQ2_XXS GGUF in llama.cpp; longer contexts, images and video need more. This system RAM is 64 GB. That is Unsloth's planning estimate, not a measured run on this hardware. Nemotron 3.5 Super VL 120B-A12B: We found no source covering AMD or Intel GPUs for this model, so only system RAM is counted. Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 64 GB of RAM = 64 GB. Longer contexts and video need more. On 1 of 3 cards: Radeon AI PRO R9700 – Not covered On 1 of 3 cards: Tesla V100 |
No answer on any of these cards: Mistral Large 4, Kolibri-1 (official FP8 route), GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar
| Model (checker) | With 64 GB RAM, card by card |
|---|---|
| Mistral Large 4 | … Weights not released smallest estimate ~210 GB vs 96 GB of memory (VRAM + RAM, or a Mac's unified memory) On all 3 cards |
| Kolibri-1 (official FP8 route) | – Not covered On all 3 cards |
| GLM-5.3-Flash (Unsloth GGUF) | ✕ Does not fit Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. On 1 of 3 cards: GeForce RTX 5090 ✕ Does not fit Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 64 GB of RAM (this card is not counted: Unsloth names CUDA, CPU-only and Metal builds, no AMD or Intel GPU build). On 1 of 3 cards: Radeon AI PRO R9700 – Not covered On 1 of 3 cards: Tesla V100 |
| DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar | – Not covered On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 – Best size: not covered DeepSeek V4.1 Flash September release: not covered; DeepSeek V4.1 Flash 1017 update: not covered On 1 of 3 cards: GeForce RTX 5090 |
Image models
| Model (checker) | With 64 GB RAM, card by card |
|---|---|
| Qwen-Image-2.1 (image generation) | ! Planning estimate only Within the all-components-resident planning budget; allocation unverified 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. On 1 of 3 cards: GeForce RTX 5090 – Not covered On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 |
| FLUX 3 Image Klein (int8, ComfyUI) | ! A published route or recipe matches, with caveats Matches BFL's 8 GB guidance: the int8 files in ComfyUI 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. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). On 1 of 3 cards: GeForce RTX 5090 – Not covered On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 |
Video models
| Model (checker) | With 64 GB RAM, card by card |
|---|---|
| MiniMax H3 (video + audio) | ! A published route or recipe matches, with caveats Hardware memory matches a published video recipe; exact clip peak unverified 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. Matching recipes: SGLang, one 24 GB card (measured): one card, DiT and text encoder streamed layer by layer; GPU peak about 18 GB; RTX 4090 D: 1344×768, 107 frames, 20 steps in 405.6 s at BF16, 303.3 s with convrot_int8 (host RAM not stated for this run; we require the 32 GB of the cookbook's smallest recipe (RigCheck assumption); a 4090 on a 32 GB host has not been measured) | 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 On 1 of 3 cards: GeForce RTX 5090 – Not covered On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700 |
Quick setup, as in the home page finder: one card, a 32K context and no image input for the llama.cpp capacity estimates (Nemotron 3.5, GLM-5.3-Flash and DeepSeek V4.1 Flash are matched against published recipes, memory tables and routes at their own context instead (NVIDIA's 64K / 256K recipes; Unsloth's GLM table states no context length; DwarfStar's routes per DeepSeek checkpoint (documented for the September weights, experimental for 1017), not llama.cpp).), the better of Windows and Linux for Strata, an NVMe SSD, 64 GB of system RAM; for the image and video rows, 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. "With caveats" means part of the model sits in system RAM (slower), an unmeasured peak, or a route the checker marks conditional. Cards with the same answer, word for word, share a line; "RTX 3070 / 3070 Ti" is two cards the checkers treat alike. There is no speed line, because speeds differ between cards with the same memory; where a checker's own conditions mention a speed, they name the hardware it was measured on. Each card's page and the full checkers show measured speeds, labeled. Pick your card in a checker for your exact setup.
What system RAM changes
The checkers' own answers for the same cards with 32 GB instead of 64 GB of system RAM, and, for 64 GB, the smallest larger RAM size for which a checker gives a better answer. Parts of a large model that do not fit on the card are kept in system RAM, so RAM matters as much as VRAM for the big mixture-of-experts models.
With 32 GB of RAM
- Qwen3.8-Flash-Next in the Strata engine. With 32 GB: ! Runs, with caveats; IQ2_XS; Runs in low-RAM mode: the 24 GB+ card holds part of the experts and the rest are copied from a file on the SSD into RAM at start. Measured with 64 GB of RAM; with 32 GB your PC runs it in low-RAM mode, which is slower. For code, the Coder (IQ1_M) runs here with its experts copied into RAM at start; Strata: it is weaker outside code and in languages other than English. Strata's installer pre-selects IQ3_S for 64 GB of RAM + VRAM, which would read experts from the SSD in low-RAM mode here (slower); RigCheck picks IQ2_XS, whose experts Strata copies into RAM at start on this PC. With 64 GB: ✓ Runs; IQ3_S; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. On 1 of 3 cards: GeForce RTX 5090.
- Qwen3.8-Flash-Next in the Strata engine. With 32 GB: ! Experimental path; IQ2_XS; Runs in low-RAM mode: the 24 GB+ card holds part of the experts and the rest are copied from a file on the SSD into RAM at start. For code, the Coder (IQ1_M) runs here with its experts copied into RAM at start; Strata: it is weaker outside code and in languages other than English. Strata's installer pre-selects IQ3_S for 64 GB of RAM + VRAM, which would read experts from the SSD in low-RAM mode here (slower); RigCheck picks IQ2_XS, whose experts Strata copies into RAM at start on this PC. Pascal / Volta cards use Strata's separate CUDA 12 engine (the cuda12 app on Windows or Linux, or a Linux build), covered by the community "Older GPUs" guide. With 64 GB: ! Experimental path; IQ3_S; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. Pascal / Volta cards use Strata's separate CUDA 12 engine (the cuda12 app on Windows or Linux, or a Linux build), covered by the community "Older GPUs" guide. On 1 of 3 cards: Tesla V100.
- Qwen3.8-Flash-Next in the Strata engine. With 32 GB: ! Runs, with caveats; IQ2_XS; Runs in low-RAM mode: the 24 GB+ card holds part of the experts and the rest are copied from a file on the SSD into RAM at start. For code, the Coder (IQ1_M) runs here with its experts copied into RAM at start; Strata: it is weaker outside code and in languages other than English. Strata's installer pre-selects IQ3_S for 64 GB of RAM + VRAM, which would read experts from the SSD in low-RAM mode here (slower); RigCheck picks IQ2_XS, whose experts Strata copies into RAM at start on this PC. With 64 GB: ✓ Runs; IQ3_S; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. On 1 of 3 cards: Radeon AI PRO R9700.
- Qwen3.8-Flash-Next in llama.cpp. With 32 GB: ✕ Does not fit; Even the smallest file (UD-IQ1_S, 72.5 GB) needs about 76.1 GB. With 64 GB: ! Fits with caveats (capacity estimate); UD-Q2_K_XL · GPU and system RAM (53.1 GB in RAM); Flash-Next uses combined memory totals; per-card layer placement is not modeled. Close fits are uncertain until runtime allocation is checked.; RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090.
- Mistral Medium 3.5 in llama.cpp. With 32 GB: ! Fits with caveats (capacity estimate); UD-IQ2_XXS · GPU and system RAM (52 of 88 layers on the GPU): a dense model runs slowly from RAM; RigCheck capacity estimate, not a measured setup. With 64 GB: ! Fits with caveats (capacity estimate); IQ4_XS · GPU and system RAM (28 of 88 layers on the GPU): a dense model runs slowly from RAM; RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: GeForce RTX 5090.
- Qwen3.8-27B in llama.cpp. With 32 GB: ! Fits with caveats (capacity estimate); UD-Q4_K_XL · system RAM, CPU only; RigCheck capacity estimate, not a measured setup. With 64 GB: ! Fits with caveats (capacity estimate); Q8_0 · system RAM, CPU only; RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100.
- Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp. With 32 GB: ! Best size: fits with caveats (capacity estimate); Gemma 4 31B: fits with caveats (capacity estimate) (IQ4_XS, system RAM, CPU only); Gemma 4 26B-A4B: fits with caveats (capacity estimate) (UD-Q4_K_XL, system RAM, CPU only); Gemma 4 12B: fits with caveats (capacity estimate) (Q8_0, system RAM, CPU only); Gemma 4 E4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E2B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); RigCheck capacity estimate, not a measured setup. With 64 GB: ! Best size: fits with caveats (capacity estimate); Gemma 4 31B: fits with caveats (capacity estimate) (Q8_0, system RAM, CPU only); Gemma 4 26B-A4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 12B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E2B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100.
- Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B). With 32 GB: ! Best size: a published route or recipe matches, with caveats; Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 64 GB budget: UD-IQ1_M (46.1 GB) or UD-IQ2_XXS (51.1 GB)); Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 32 GB of RAM = 64 GB. Longer contexts and video need more. With 64 GB: ! Best size: a published route or recipe matches, with caveats; Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 96 GB budget: UD-Q3_K_XL (65.8 GB) or NVFP4 (79 GB)); Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. Longer contexts and video need more. On 1 of 3 cards: GeForce RTX 5090.
- Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B). With 32 GB: ! Best size: a published route or recipe matches, with caveats; Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches Unsloth's short-context budget: about 32 GB of RAM or unified memory); Nemotron 3.5 Super VL 120B-A12B: does not fit; Nemotron 3.5 Bolt 74B-A9B: We found no source covering AMD GPUs for this model, so only system RAM is counted here. Unsloth budgets around 32 GB of RAM or unified memory for short-context use (48–64 GB for more headroom) with its 22 GB IQ2_XXS GGUF in llama.cpp; longer contexts, images and video need more. This system RAM is 32 GB. That is Unsloth's planning estimate, not a measured run on this hardware. Nemotron 3.5 Super VL 120B-A12B: We found no source covering AMD or Intel GPUs for this model, so only system RAM is counted. Unsloth plans at least 64 GB of RAM or unified memory for its smallest GGUF (UD-IQ1_M, 46.1 GB); llama.cpp can split the model across RAM and VRAM. This setup counts 32 GB of RAM = 32 GB. With 64 GB: ! Best size: a published route or recipe matches, with caveats; Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches Unsloth's short-context budget: about 32 GB of RAM or unified memory); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 64 GB budget: UD-IQ1_M (46.1 GB) or UD-IQ2_XXS (51.1 GB)); Nemotron 3.5 Bolt 74B-A9B: We found no source covering AMD GPUs for this model, so only system RAM is counted here. Unsloth budgets around 32 GB of RAM or unified memory for short-context use (48–64 GB for more headroom) with its 22 GB IQ2_XXS GGUF in llama.cpp; longer contexts, images and video need more. This system RAM is 64 GB. That is Unsloth's planning estimate, not a measured run on this hardware. Nemotron 3.5 Super VL 120B-A12B: We found no source covering AMD or Intel GPUs for this model, so only system RAM is counted. Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 64 GB of RAM = 64 GB. Longer contexts and video need more. On 1 of 3 cards: Radeon AI PRO R9700.
With more than 64 GB of RAM
- Qwen3.8-Flash-Next in llama.cpp: with 96 GB of RAM, fits with caveats (capacity estimate) (UD-IQ3_XXS, system RAM, CPU only). On 2 of 3 cards: Tesla V100, Radeon AI PRO R9700.
The 32 GB cards on this page
3 cards in 3 finder entries; an entry such as "RTX 3070 / 3070 Ti" is two cards the checkers treat alike. Each card's memory as its vendor states it (read 2026-10-18):
- GeForce RTX 5090: 32 GB GDDR7 (NVIDIA GeForce RTX 5090 specifications)
- Tesla V100: GPU Memory 32GB /16GB HBM2 (NVIDIA Tesla V100 datasheet (PDF))
- Radeon AI PRO R9700: 32 GB GDDR6 (AMD Radeon AI PRO R9700 specifications)
A Mac with 32 GB of unified memory is a different machine: its memory is shared by the graphics and the processor, see its own page.
FAQ
Can a 32 GB graphics card run Qwen3.8-Flash-Next in the Strata engine?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on all 3 32 GB cards in the finder. On 2 of 3 cards: GeForce RTX 5090, Radeon AI PRO R9700: Yes: IQ3_S; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. On 1 of 3 cards: Tesla V100: Only on an experimental path: IQ3_S; Fits by the docs' rule (experts + ~10 GB), with little room to spare: close the browser while it runs. Pascal / Volta cards use Strata's separate CUDA 12 engine (the cuda12 app on Windows or Linux, or a Linux build), covered by the community "Older GPUs" guide. The full checker takes your exact setup (RAM, disk, operating system, processor) and gives the install steps.
Can a 32 GB graphics card run Qwen3.8-Flash-Next in llama.cpp?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 1 of the 3 32 GB cards in the finder: on 1 of them the model is placed wholly or partly on the card. On 1 of 3 cards: GeForce RTX 5090: With caveats, by RigCheck's capacity estimate: UD-Q2_K_XL · GPU and system RAM (53.1 GB in RAM); Flash-Next uses combined memory totals; per-card layer placement is not modeled. Close fits are uncertain until runtime allocation is checked; RigCheck capacity estimate, not a measured setup. No answer on: Tesla V100, Radeon AI PRO R9700. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can a 32 GB graphics card run Mistral Medium 3.5 in llama.cpp?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 1 of the 3 32 GB cards in the finder: on 1 of them the model is placed wholly or partly on the card. On 1 of 3 cards: GeForce RTX 5090: With caveats, by RigCheck's capacity estimate: IQ4_XS · GPU and system RAM (28 of 88 layers on the GPU): a dense model runs slowly from RAM; RigCheck capacity estimate, not a measured setup. No answer on: Tesla V100, Radeon AI PRO R9700. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can a 32 GB graphics card run Qwen3.8-27B in llama.cpp?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on all 3 32 GB cards in the finder: on 2 of them the model is placed wholly or partly on the card, on 1 it runs on the processor in system RAM alone. On 1 of 3 cards: GeForce RTX 5090: Yes, by RigCheck's capacity estimate: UD-Q6_K_XL · GPU memory (input embeddings in system RAM); RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Radeon AI PRO R9700: Yes, by RigCheck's capacity estimate: UD-Q6_K_XL · GPU memory (input embeddings in system RAM); On this Radeon card this is a capacity estimate only: we found no source for this model with llama.cpp's own AMD builds. Ollama lists this card for its ROCm support, but its default package (ollama run qwen3.8:27b, 18 GB) is a different file from the one checked here, so compare its size with your memory first; RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100: With caveats, by RigCheck's capacity estimate: Q8_0 · system RAM, CPU only; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can a 32 GB graphics card run Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on all 3 32 GB cards in the finder: on 2 of them the model is placed wholly or partly on the card, on 1 it runs on the processor in system RAM alone. On 1 of 3 cards: GeForce RTX 5090: Yes, by RigCheck's capacity estimate: Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Radeon AI PRO R9700: Yes, by RigCheck's capacity estimate: Gemma 4 31B: fits (capacity estimate) (Q5_K_M, GPU memory (input embeddings in system RAM)); Gemma 4 26B-A4B: fits (capacity estimate) (Q8_0, GPU memory (input embeddings in system RAM)); Gemma 4 12B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E4B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); Gemma 4 E2B: fits (capacity estimate) (BF16, GPU memory (input embeddings in system RAM)); On this Radeon card these are capacity estimates only: we found no source for Gemma 4 with llama.cpp's own AMD builds. Ollama lists this card for its ROCm support, but its default packages (gemma4:31b 19–20 GB, gemma4:26b 16–19 GB, gemma4:12b 7.7–8.0 GB, gemma4:e4b 6.6–9.5 GB, gemma4:e2b 4.6–7.5 GB) are different files from the ones checked here; RigCheck capacity estimate, not a measured setup. On 1 of 3 cards: Tesla V100: With caveats, by RigCheck's capacity estimate: Gemma 4 31B: fits with caveats (capacity estimate) (Q8_0, system RAM, CPU only); Gemma 4 26B-A4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 12B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E4B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); Gemma 4 E2B: fits with caveats (capacity estimate) (BF16, system RAM, CPU only); RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.
Can a 32 GB graphics card run Qwen-Image-2.1 (image generation)?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 1 of the 3 32 GB cards in the finder. On 1 of 3 cards: GeForce RTX 5090: Only as a planning estimate: the checker's uncalibrated memory budget, offload peak unverified: Within the all-components-resident planning budget; allocation unverified; 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. No answer on: Tesla V100, Radeon AI PRO R9700. The full checker lets you change the task, resolution, references and offload (batch is fixed at one), and lists the files.
Can a 32 GB graphics card run MiniMax H3 (video + audio)?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 1 of the 3 32 GB cards in the finder. On 1 of 3 cards: GeForce RTX 5090: A published route or recipe matches, with caveats: Hardware memory matches a published video recipe; exact clip peak unverified; 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; Matching recipes: SGLang, one 24 GB card (measured): one card, DiT and text encoder streamed layer by layer; GPU peak about 18 GB; RTX 4090 D: 1344×768, 107 frames, 20 steps in 405.6 s at BF16, 303.3 s with convrot_int8 (host RAM not stated for this run; we require the 32 GB of the cookbook's smallest recipe (RigCheck assumption); a 4090 on a 32 GB host has not been measured) | 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. No answer on: Tesla V100, Radeon AI PRO R9700. The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Can a 32 GB graphics card run Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B)?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 2 of the 3 32 GB cards in the finder. On 1 of 3 cards: GeForce RTX 5090: A published route or recipe matches, with caveats: Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches NVIDIA's 28GB llama.cpp recipe: 64K context, q8_0 KV cache, one request at a time); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 96 GB budget: UD-Q3_K_XL (65.8 GB) or NVFP4 (79 GB)); Nemotron 3.5 Bolt 74B-A9B: NVIDIA's 256K-context recipe needs an RTX Spark 128 GB or a 48 GB+ discrete GPU; this hardware gets the 65,536-token recipe. Build the latest llama.cpp from source first. No speed is published for this machine; NVIDIA's 84 tokens per second is from a 128 GB RTX Spark. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. Longer contexts and video need more. On 1 of 3 cards: Radeon AI PRO R9700: A published route or recipe matches, with caveats: Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches Unsloth's short-context budget: about 32 GB of RAM or unified memory); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 64 GB budget: UD-IQ1_M (46.1 GB) or UD-IQ2_XXS (51.1 GB)); Nemotron 3.5 Bolt 74B-A9B: We found no source covering AMD GPUs for this model, so only system RAM is counted here. Unsloth budgets around 32 GB of RAM or unified memory for short-context use (48–64 GB for more headroom) with its 22 GB IQ2_XXS GGUF in llama.cpp; longer contexts, images and video need more. This system RAM is 64 GB. That is Unsloth's planning estimate, not a measured run on this hardware. Nemotron 3.5 Super VL 120B-A12B: We found no source covering AMD or Intel GPUs for this model, so only system RAM is counted. Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 64 GB of RAM = 64 GB. Longer contexts and video need more. No answer on: Tesla V100. The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Can a 32 GB graphics card run FLUX 3 Image Klein (int8, ComfyUI)?
In RigCheck's quick setup with 64 GB of system RAM, the checker has an answer on 1 of the 3 32 GB cards in the finder. On 1 of 3 cards: GeForce RTX 5090: A published route or recipe matches, with caveats: Matches BFL's 8 GB guidance: the int8 files in ComfyUI; 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. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). No answer on: Tesla V100, Radeon AI PRO R9700. The full checker lets you change the route or recipe settings and lists the files and the published recipe.
Why do 32 GB cards get different answers?
Each checker judges the card itself, with its own support list and the routes its sources name, not only the memory size. On this page that happens for Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen3.8-27B in llama.cpp, Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp, Qwen-Image-2.1 (image generation), MiniMax H3 (video + audio), Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B), FLUX 3 Image Klein (int8, ComfyUI), GLM-5.3-Flash (Unsloth GGUF), DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar; the table lists each card's answer. Speeds also differ between cards with the same memory, so this page has no speed line: each card's own page and the full checkers show measured speeds, labeled.
How does system RAM change what a 32 GB card can run?
With 32 GB instead of 64 GB, the answer changes on some or all of these cards for 6 models: Qwen3.8-Flash-Next in the Strata engine, Qwen3.8-Flash-Next in llama.cpp, Mistral Medium 3.5 in llama.cpp, Qwen3.8-27B in llama.cpp, Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp, Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B). With more than 64 GB, a checker gives a better answer for 1 model: Qwen3.8-Flash-Next in llama.cpp. See What system RAM changes for the cards and the answers.
Other memory sizes
8 GB · 12 GB · 16 GB · 24 GB · All VRAM sizes · Graphics cards · Macs · Any other machine
Sources
- DeepSeek V4.1 Flash (September release, 1017 update) in DwarfStar: the checker, its sources and assumptions
- FLUX 3 Image Klein (int8, ComfyUI): the checker, its sources and assumptions
- Gemma 4 (E2B, E4B, 12B, 26B-A4B, 31B) in llama.cpp: the checker, its sources and assumptions
- GLM-5.3-Flash (Unsloth GGUF): the checker, its sources and assumptions
- Kolibri-1 (official FP8 route): the checker, its sources and assumptions
- MiniMax H3 (video + audio): the checker, its sources and assumptions
- Mistral Large 4: the checker, its sources and assumptions
- Mistral Medium 3.5 in llama.cpp: the checker, its sources and assumptions
- Nemotron 3.5 (Bolt 74B-A9B, Super VL 120B-A12B): the checker, its sources and assumptions
- Qwen3.8-27B in llama.cpp: the checker, its sources and assumptions
- Qwen3.8-Flash-Next in llama.cpp: the checker, its sources and assumptions
- Qwen-Image-2.1 (image generation): the checker, its sources and assumptions
- Qwen3.8-Flash-Next in the Strata engine: the checker, its sources and assumptions
- NVIDIA Tesla V100 datasheet (PDF): card memory
- AMD Radeon AI PRO R9700 specifications: card memory
- NVIDIA GeForce RTX 5090 specifications: card memory