AI models on a 12 GB graphics card
Open chat models judged for a PC card with 12 GB of video memory: which fit with every part on the card, the size to download, and what each one takes.
3043 of the 3940 chat models people download as GGUF fit a 12 GB card whole. The best-rated is Gemma 3 12B at Q6_K, which takes 10 GB. The largest rated model that fits is Internlm2 5 20B (20B).
Cards with 12 GB: GeForce RTX 4070, GeForce RTX 3060, GeForce RTX 5070, GeForce RTX 5070 Ti Laptop, GeForce RTX 4070 Ti, GeForce RTX 4070 Super, GeForce RTX 4080 Laptop, GeForce RTX 3080 Ti, GeForce RTX 3080 12GB, Radeon RX 7700 XT, Radeon RX 6700 XT, Arc B580. The same answer holds for any card with 12 GB of its own memory.
The best-rated models that fit
The 25 with the highest LM Arena rating (data, CC BY 4.0) that fit at 4 bits a weight or more, each at the most bits that fits. Below 4 bits a model's rating overstates it. Each model's page hands it to LM Studio, Jan, Ollama and llama.cpp.
| Model | Size | Fit | Takes | Longest chat | Arena | Arena coding |
|---|---|---|---|---|---|---|
| Gemma 3 12BGoogle · 12B | Q6_K9.7 GB · unsloth | Fits with room | 10 GB | 8K | 1342 | 1316 |
| Gemma 3n E4BGoogle · 6.9B | Q8_07.4 GB · unsloth | Fits with room | 7.7 GB | 32K | 1318 | 1307 |
| GPT-OSS 20BOpenAI · 21B, 4.2B active | Q4_K_S12 GB · unsloth | Fits | 12 GB | 8K | 1317 | 1368 |
| Granite 4.1 8BIBM · 8.8B | Q8_09.3 GB · ibm-granite | Fits with room | 10 GB | 8K | 1304 | 1353 |
| Gemma 3 4BGoogle · 3.9B | Q6_K_L3.4 GB · bartowski | Fits with room | 5.1 GB | 128K | 1303 | 1273 |
| Granite 4.2 8BIBM · 8.8B | Q8_09.3 GB · ibm-granite | Fits with room | 10 GB | 8K | 1290 | 1370 |
| Granite 4.2 3BIBM · 3.7B | Q8_03.9 GB · ibm-granite | Fits with room | 10 GB | 128K | 1290 | 1378 |
| Gemma 2 9BGoogle · 9.2B | Q6_K7.6 GB · bartowski | Fits with room | 9.3 GB | 8K | 1267 | 1270 |
| Phi 4Microsoft · 15B | Q4_K9.1 GB · microsoft | Fits with room | 10 GB | 8K | 1256 | 1306 |
| Aya Expanse 8BCohere · 8B | Q8_08.5 GB · bartowski | Fits with room | 9.3 GB | 8K | 1223 | 1255 |
| Llama 3.1 Tulu 3 8BAi2 · 8B | Q8_08.5 GB · lmstudio-community | Fits with room | 9.9 GB | 16K | 1221 | |
| Llama 3.1 8BMeta · 8B | UD-Q6_K_XL7.3 GB · unsloth | Fits with room | 9.9 GB | 32K | 1211 | 1259 |
| Granite 3.1 8BIBM · 8.2B | Q8_08.7 GB · bartowski | Fits with room | 9.6 GB | 8K | 1209 | |
| Gemma 2 2BGoogle · 2.6B | Q8_02.8 GB · bartowski | Fits with room | 3.3 GB | 8K | 1200 | 1191 |
| Phi 3 Medium 4KMicrosoft · 14B | Q5_K_S9.6 GB · bartowski | Fits with room | 10 GB | 4K | 1198 | 1231 |
| Internlm2 5 20BShanghai AI Lab · 20B | IQ4_XS11 GB · bartowski | Fits | 12 GB | 8K | 1191 | 1249 |
| Granite 3.1 2BIBM · 2.5B | Q8_02.7 GB · bartowski | Fits with room | 9 GB | 128K | 1179 | 1249 |
| Llama 3.2 3BMeta · 3.2B | Q8_03.4 GB · unsloth | Fits with room | 7.9 GB | 64K | 1167 | 1176 |
| SOLAR 10.7B V1.0Upstage · 11B | Q6_K8.8 GB · TheBloke | Fits with room | 9.4 GB | 4K | 1152 | |
| Mistral 7B V0.2Mistral AI · 7.2B | Q8_07.7 GB · TheBloke | Fits with room | 10 GB | 32K | 1149 | 1186 |
| Qwen 14BAlibaba Qwen · 14B | Q3_K_M7.7 GB · mradermacher | Fits | 12 GB | 8K | 1139 | 1198 |
| Phi 3 Mini 128KMicrosoft · 3.8B | Q8_04.1 GB · QuantFactory | Fits with room | 7.9 GB | 16K | 1130 | 1155 |
| Phi 3 Mini 4KMicrosoft · 3.8B | Q8_04.1 GB · bartowski | Fits with room | 5.1 GB | 4K | 1128 | 1187 |
| SmolLM2 1.7BHugging Face · 1.7B | Q8_01.8 GB · bartowski | Fits with room | 2.9 GB | 8K | 1115 | |
| Llama 3.2 1BMeta · 1.2B | Q8_01.3 GB · unsloth | Fits with room | 3.9 GB | 128K | 1111 | 1149 |
The chat models Outmake picks
16 of these 28 fit. Their memory is what llama.cpp's own calculator reported for each size, not an estimate.
| Model | Size | Fit | Takes |
|---|---|---|---|
| K2-Horizon 7BInstitute of Foundation Models · 9B | Q6_K7.8 GB | Fits with room | 10 GB |
| K2-Horizon 3.7BInstitute of Foundation Models · 5.1B | Q6_K4.4 GB | Fits with room | 7.0 GB |
| Gemma 4 12BGoogle · 12B | Q4_K_M7.3 GB | Fits with room | 9.2 GB |
| Qwen3.5 9BAlibaba Qwen · 9B | Q6_K8.4 GB | Fits with room | 8.7 GB |
| Gemma 4 E4BGoogle · 8B | UD-Q5_K_XL7.6 GB | Fits with room | 7.6 GB |
| Gemma 4 E2BGoogle · 5.1B | Q6_K5.5 GB | Fits with room | 5.5 GB |
| Qwen3.5 4BAlibaba Qwen · 4B | Q6_K4.2 GB | Fits with room | 7.3 GB |
| GPT-OSS 20BOpenAI · 21B, 3.6B active | Q3_K_M12 GB | Fits | 12 GB |
| Qwen3.5 2BAlibaba Qwen · 2B | Q6_K2.2 GB | Fits with room | 3.6 GB |
| Qwen3.5 0.8BAlibaba Qwen · 0.8B | Q6_K0.8 GB | Fits with room | 2.4 GB |
| Qwen3.8 27BAlibaba Qwen · 28B | UD-Q2_K_XL11 GB | Fits | 12 GB |
| Gemma 4 31BGoogle · 31B | UD-IQ2_XXS9.7 GB | Fits | 11 GB |
| K2-Horizon 36BInstitute of Foundation Models · 37B, 4B active | IQ2_XXS11 GB | Fits | 12 GB |
| Qwen3.6 35BAlibaba Qwen · 36B, 3B active | UD-IQ2_XXS12 GB | Fits | 12 GB |
| GLM 4.7 FlashZ.ai · 31B, 3B active | Q2_K_L11 GB | Fits | 12 GB |
| Devstral Small 2 24BMistral AI · 24B | UD-IQ3_XXS10 GB | Fits | 11 GB |
| Gemma 4 26BGoogle · 26B, 3.8B active | UD-IQ2_M11 GB | Tight fit | 12 GB |
| Qwen3.5 122BAlibaba Qwen · 122B, 10B active | UD-Q3_K_XL58 GB | Bigger than the card | 62 GB |
| GPT-OSS 120BOpenAI · 117B, 5.1B active | UD-Q4_K_XL63 GB | Bigger than the card | 65 GB |
| Qwen3.8 Flash NextAlibaba Qwen · 177B, 6B active | UD-Q3_K_XL91 GB | Bigger than the card | 91 GB |
| MiniMax M2.7MiniMax · 229B, 10B active | UD-Q3_K_XL102 GB | Bigger than the card | 120 GB |
| Qwen3.5 397BAlibaba Qwen · 403B, 17B active | UD-IQ2_M124 GB | Bigger than the card | 129 GB |
| DeepSeek V4 FlashDeepSeek · 284B, 13B active | UD-Q3_K_XL128 GB | Bigger than the card | 129 GB |
| GLM 5.3Z.AI · 754B, 40B active | UD-Q2_K_XL254 GB | Bigger than the card | 265 GB |
| Kimi K2.6Moonshot AI · 1T, 32B active | UD-Q2_K_XL341 GB | Bigger than the card | 351 GB |
| Kimi K3Moonshot AI · 2.8T, 104B active | UD-IQ2_XXS712 GB | Bigger than the card | 716 GB |
| Qwen3.8 2.4TAlibaba Qwen · 2.4T, 95B active | UD-IQ2_XS731 GB | Bigger than the card | 736 GB |
| DeepSeek V4 ProDeepSeek · 1.6T, 49B active | UD-Q4_K_XL850 GB | Bigger than the card | 851 GB |
How this is worked out
A model fits when its weights, its memory of the chat so far and the engine's own share come to 11 GB or less: 92% of the card, with the rest left for the screen. "Fits with room" is under 80%. "Tight fit" is anything up to half a gigabyte short of the whole card: close other programs that draw with the card first.
"Bigger than the card" does not mean it won't run. llama.cpp and the apps built on it put what doesn't fit in ordinary memory and the model answers slower, often several times slower. So this page lists what fits whole.
There are no speeds here. Outmake's speed estimates are fitted to measurements on Apple chips, and no PC has been measured yet, so a number would be a guess. Memory was checked against the catalog of November 4, 2026; which cards have 12 GB against each maker's specification page on October 13, 2026.