RunsOnMy ModelsGPUs & MacsBy VRAMBest GPUCan I run…?OllamaStart hereAPI pricesVRAM calculatorNew

What AI models run on the RTX 4070 Ti?

The RTX 4070 Ti has 12 GB of VRAM (504 GB/s). About 11.4 GB is usable for a model. 173 of 298 open models we track fit at 8K context.

173models fit (8K ctx)
154still fit at 32K ctx
31Blargest that fits
504 GB/smemory bandwidth

This card has 12 GB of VRAM. See every model that fits in 12 GB VRAM. Other 12 GB hardware: RTX 3060 12GB, RTX 3080 12GB, RTX 4070, RTX 4070 Super, RTX 5070, RX 7700 XT, Arc B580.

Using a specific tool? Ollama models that run on the RTX 4070 Ti · LM Studio · llama.cpp offload

Models that fit, largest first

ModelParamsBest quant (8K)Needs32K contextMax tok/s (est.)Details
GLM 4.7 Flash31B (4.6B act.)Q2_K11.4 GBno309details
Muse Glimmer 30B30BQ2_K11.4 GBno45details
Raon 1 Mini Preview29B (4.3B act.)Q2_K11.2 GBno311details
Arete 27B27BQ2_K11.2 GBno48details
MiniMax H3 Image PE27BQ2_K11.3 GBno47details
Trinity Mini26B (2.0B act.)Q2_K9.8 GBQ2_K711details
GLiDE 27B26BQ2_K10.5 GBno52details
DiffusionGemma 26B A4B IT26B (4.0B act.)Q2_K10.4 GBQ2_K336details
LFM2.5 24B A2B24B (2.0B act.)Q2_K9.2 GBQ2_K672details
ERNIE 4.5 21B A3B PT22BQ2_K8.6 GBQ2_K62details
ERNIE 4.5 21B A3B Thinking22BQ2_K8.6 GBQ2_K62details
Kimi VL A3B Thinking 250616B (3.0B act.)Q3_K_M8.4 GBQ3_K_M344details
Kimi VL A3B Thinking16B (3.0B act.)Q3_K_M8.4 GBQ3_K_M344details
Kimi VL A3B Instruct16B (3.0B act.)Q3_K_M8.7 GBQ3_K_M332details
DeepSeek MoE 16B Base16B (2.8B act.)Q3_K_M10.2 GBno344details
LLaDA2.1 Mini16B (1.4B act.)Q3_K_M8.4 GBQ3_K_M720details
Moonlight 16B A3B Instruct16B (3.0B act.)Q3_K_M8.7 GBQ3_K_M324details
DeepSeek Coder V2 Lite Instruct16B (2.7B act.)Q3_K_M8.0 GBQ3_K_M388details
DeepSeek V2 Lite16B (2.7B act.)Q4_K_M9.8 GBQ3_K_M310details
DeepSeek V2 Lite Chat16B (2.7B act.)Q3_K_M8.4 GBQ3_K_M366details
Qwen2.5 Coder 14B Instruct15BQ3_K_M9.0 GBno69details
DeepSeek R1 Distill Qwen 14B15BQ3_K_M9.0 GBno69details
Qwen3 14B15BQ3_K_M8.7 GBQ2_K70details
Phi 415BQ3_K_M9.0 GBno68details
Gemma 3 12B IT12BQ5_K_M9.6 GBQ4_K_M60details
Llama Guard 4 12B12BQ5_K_M10.2 GBQ2_K59details
KORMo Web 11B11BQ5_K_M9.3 GBQ2_K63details
Llama 3.2 11B Vision Instruct11BQ6_K10.0 GBQ2_K58details
Llama 3.2 11B Vision11BQ6_K10.0 GBQ2_K58details
GLM 4.6V Flash10BQ6_K8.7 GBQ6_K61details
GLM 4.1V 9B Thinking10BQ6_K8.7 GBQ6_K61details
Step3 VL 10B10BQ8_09.9 GBQ4_K_M58details
SenseNova Loop 9B Preview9.7BQ8_010.0 GBQ6_K53details
Qwen3.5 9B9.7BQ8_010.3 GBQ6_K51details
Ornith 1.5 9B9.7BQ8_010.0 GBQ6_K53details
SenseNova Loop 9B9.4BQ8_010.0 GBQ6_K53details
MiMo V2.6 Distill Qwen 9B9.4BQ8_010.0 GBQ6_K53details
GLM 4 9B9.4BQ6_K8.7 GBQ6_K61details
Apertus v1.5 8B8.9BQ6_K8.4 GBQ4_K_M69details
InternLM3 8B Instruct8.8BQ8_09.7 GBQ6_K54details
Granite 4.2 8B8.8BQ6_K8.6 GBQ3_K_M70details
Granite 4.1 8B8.8BQ6_K8.6 GBQ3_K_M70details
Granite 4.1 8B Llmon Preview8.8BQ6_K8.6 GBQ3_K_M70details
Qwen3 VL 8B Instruct8.8BQ8_09.9 GBQ4_K_M58details
Molmo2 8B8.7BQ8_09.9 GBQ4_K_M58details
LFM2.5 8B A1B8.5B (1.0B act.)Q8_09.1 GBQ8_0474details
Qwen2.5 VL 7B Instruct8.3BQ8_08.7 GBQ8_062details
UI TARS 1.5 7B8.3BQ8_08.7 GBQ8_062details
olmOCR 2 7B 10258.3BQ8_08.7 GBQ8_062details
olmOCR 7B 0225 Preview8.3BQ8_08.7 GBQ8_062details
UI TARS 7B DPO8.3BQ8_08.7 GBQ8_062details
Qwen3 8B8.2BQ8_09.9 GBQ4_K_M58details
DeepSeek R1 0528 Qwen3 8B8.2BQ8_09.9 GBQ4_K_M58details
Granite 3.3 8B Instruct8.2BQ8_09.8 GBQ4_K_M58details
Granite 3.1 8B Instruct8.2BQ8_09.8 GBQ4_K_M58details
Granite 3.2 8B Instruct8.2BQ8_09.8 GBQ4_K_M58details
Apertus 8B Instruct 25098.1BQ8_09.6 GBQ5_K_M59details
Llama 3.1 8B Instruct8.0BQ8_09.6 GBQ5_K_M59details
Llama 3.1 8B8.0BQ8_09.6 GBQ5_K_M59details
Hermes 3 Llama 3.1 8B8.0BQ8_09.6 GBQ5_K_M59details
Llama Guard 3 8B8.0BQ8_09.6 GBQ5_K_M59details
DeepSeek R1 Distill Llama 8B8.0BQ8_09.6 GBQ5_K_M59details
Meta Llama 3 8B Instruct8.0BQ8_09.6 GBQ5_K_M59details
Meta Llama Guard 2 8B8.0BQ8_09.6 GBQ5_K_M59details
Meta Llama 3.1 8B Instruct8.0BQ8_09.6 GBQ5_K_M59details
Aya Expanse 8B8.0BQ8_09.8 GBQ4_K_M59details
Ling 3.0 Tiny7.9B (1.4B act.)Q8_08.7 GBQ8_0342details
MiMo 7B RL7.8BQ8_09.4 GBQ4_K_M62details
InternLM2.5 7B Chat7.7BQ8_09.2 GBQ5_K_M61details
InternLM2 Chat 7B7.7BQ8_09.2 GBQ5_K_M61details
Qwen2.5 7B Instruct7.6BQ8_08.7 GBQ8_062details
DeepSeek R1 Distill Qwen 7B7.6BQ8_08.7 GBQ8_062details
Qwen2.5 Coder 7B Instruct7.6BQ8_08.7 GBQ8_062details
Olmo 3 7B Instruct7.3BQ6_K8.7 GBQ4_K_M84details
Olmo 3 1025 7B7.3BQ6_K8.7 GBQ4_K_M84details
Olmo 3 7B Think7.3BQ6_K8.7 GBQ4_K_M84details
Mistral 7B v0.17.2BQ8_08.6 GBQ6_K65details
Mistral 7B Instruct v0.27.2BQ8_08.6 GBQ6_K65details
Mistral 7B Instruct v0.17.2BQ8_08.6 GBQ6_K65details
Falcon 7B7.2BQ8_07.7 GBQ8_066details
Falcon 7B Instruct7.2BQ8_07.4 GBQ8_069details
Granite 4.0 H Tiny6.9B (1.5B act.)Q8_07.5 GBQ8_0323details
OLMoE 1B 7B 09246.9B (1.3B act.)Q8_08.4 GBQ6_K370details
DeepSeek Coder 7B Instruct v1.56.9BQ6_K9.6 GBno89details
DeepSeek Coder 6.7B Instruct6.7BQ6_K9.6 GBno91details
DeepSeek Coder 6.7B Base6.7BQ6_K9.6 GBno91details
Llama 2 7B HF6.7BQ6_K9.6 GBno91details
Yi 6B 200K6.1BQ8_07.0 GBQ8_078details
Yi 6B6.1BQ8_07.0 GBQ8_078details
Molmo2 4B4.9BF1610.0 GBQ8_057details
WeVisOne VL 4B4.8BF1610.0 GBQ8_057details
Mage VL Cowork4.7BF1610.0 GBQ8_057details
Qianfan OCR4.7BF1610.1 GBQ8_057details
Qwen3.5 4B4.7BF169.2 GBF1658details
SingVL S1 4B4.5BF169.0 GBF1660details
LightOnOCR 3 4B4.5BF169.0 GBF1660details
Intern Decision 4B4.5BF169.0 GBF1660details
Qwen3 VL 4B Instruct4.4BF169.3 GBQ8_063details
Gemma 3 4B IT4.3BF168.4 GBF1665details
MedGemma 1.5 4B IT4.3BF168.4 GBF1665details
MedGemma 4B IT4.3BF168.4 GBF1665details
TranslateGemma 4B IT4.3BF169.2 GBF1659details
Gemma 3 4B PT4.3BF168.4 GBF1665details
Phi 3.5 Vision Instruct4.1BQ8_07.6 GBno114details
Qwen3 4B4.0BF169.3 GBQ8_063details
Qwen3 4B Instruct 25074.0BF169.3 GBQ8_063details
Granite Vision 4.1 4B4.0BF167.6 GBF1674details
Phi 4 Mini Instruct3.8BF168.9 GBQ8_066details
Phi 3 Mini 4K Instruct3.8BQ8_07.2 GBno124details
Phi 3.5 Mini Instruct3.8BQ8_07.2 GBno124details
Phi 3 Mini 128K Instruct3.8BQ8_07.2 GBno124details
Qwen2.5 VL 3B Instruct3.8BF168.0 GBF1667details
Granite 4.2 3B3.7BF168.0 GBF1669details
Granite 4.1 3B Llmon Preview3.7BF168.0 GBF1669details
Raon 1 Nano3.5BF167.0 GBF1680details
Granite 4.1 3B3.4BF167.6 GBF1674details
Granite 4.0 Micro3.4BF167.6 GBF1674details
DeepSeek OCR 23.4B (1.2B act.)F166.6 GBF16243details
DeepSeek VL2 Tiny3.4B (1.2B act.)F167.4 GBF16214details
Tiny Aya Global3.3BF167.5 GBF1675details
Unlimited OCR 23.3B (1.1B act.)F166.6 GBF16249details
Unlimited OCR3.3B (1.1B act.)F166.6 GBF16251details
DeepSeek OCR3.3B (1.1B act.)F166.6 GBF16251details
Granite 3.1 3B A800M Instruct3.3B (883M act.)F167.1 GBF16285details
Llama 3.2 3B Instruct3.2BF167.5 GBF1678details
Llama 3.2 3B3.2BF167.5 GBF1678details
Hermes 3 Llama 3.2 3B3.2BF167.5 GBF1678details
d1 3B3.1BF165.8 GBF1693details
LFM2.5 VL 3B3.1BF165.8 GBF1693details
Qwen2.5 3B Instruct3.1BF166.7 GBF1682details
SmolLM3 3B3.1BF166.9 GBF1682details
SmolLM3 3B Base3.1BF166.9 GBF1682details
Phi 22.8BF168.2 GBno91details
LFM2.5 2.6B2.7BQ8_03.4 GBQ8_0175details
Granite 3.3 2B Instruct2.5BF165.8 GBF1699details
MiniCPM5 2B2.5BF165.7 GBF16100details
WeVisOne VL 2B2.4BF165.4 GBF16124details
Cosmos Reason2 2B2.4BF165.4 GBF16124details
Qwen3.5 2B2.3BF164.6 GBF16129details
SmolVLM Instruct2.2BF165.4 GBF16139details
Qwen3 1.7B2.0BF165.4 GBF16124details
d1 2B1.9BF164.0 GBF16145details
LFM3 VL 2B1.9BF164.0 GBF16145details
DeepSeek R1 Distill Qwen 1.5B1.8BF164.2 GBF16142details
LFM3 1.7B1.7BF167.2 GBF1673details
LFM3 1.7B Base1.7BF164.0 GBF16145details
SmolLM2 1.7B Instruct1.7BF165.2 GBF16147details
SmolLM2 1.7B1.7BF165.2 GBF16147details
SmolLM 1.7B1.7BF165.2 GBF16147details
Granite 4.0 1B1.6BF164.3 GBF16154details
Qwen2.5 1.5B Instruct1.5BF163.8 GBF16163details
OLMo 2 0425 1B1.5BF164.4 GBF16169details
DeepSeek Coder 1.3B Instruct1.3BF164.5 GBF16187details
Llama 3.2 1B1.2BF163.2 GBF16203details
Llama 3.2 1B Instruct1.2BF163.2 GBF16203details
LFM2.5 1.2B Instruct1.2BF165.0 GBF16108details
LFM2 1.2B1.2BF163.0 GBF16215details
HunyuanOCR1.1BF163.5 GBF16225details
PaddleOCR VL 1.6959MF162.5 GBF16263details
Qwen3.5 0.8B873MF162.4 GBF16324details
LightOnOCR 3 0.8B853MF162.4 GBF16332details
FastVLM 0.5B759MF162.2 GBF16332details
Qwen3 0.6B752MF163.0 GBF16334details
GOT OCR2.0716MF162.8 GBF16352details
Nemotron 3.5 Parse631MF161.9 GBF16386details
GOT OCR 2.0 HF561MF162.5 GBF16450details
SmolVLM2 500M Video Instruct507MF161.6 GBF16614details
SmolVLM 500M Instruct507MF161.6 GBF16614details
Qwen2.5 0.5B Instruct494MF161.7 GBF16510details
LFM2.5 VL 450M449MF161.3 GBF16708details
SmolLM2 360M Instruct362MF161.5 GBF16695details
SmolLM2 360M362MF161.5 GBF16695details
LFM2.5 350M354MF161.9 GBF16354details

Bigger models with CPU offload

These popular models don't fit in the RTX 4070 Ti's VRAM at 4-bit, but run with llama.cpp offload, keeping part of the model in system RAM. Speed range is DDR4-3200 to DDR5-6000 (estimate). Method.

ModelParamsOffload plan
GLM 5.3 Flash321B MoEOffload Q2_K, 128 GB RAM · ~6.9–13 tok/s
gpt-oss 120B117B MoEOffload Q4_K_M, 64 GB RAM · ~13–26 tok/s
DeepSeek V4 Flash 0731304B MoEOffload Q2_K, 128 GB RAM · ~7.5–13 tok/s
DeepSeek V4 Flash291B MoEOffload Q2_K, 128 GB RAM · ~9.2–18 tok/s
GLM 4.5 Air110B MoEOffload Q4_K_M, 128 GB RAM · ~5.2–9.9 tok/s
MiniMax M2.5229B MoEOffload Q3_K_M, 128 GB RAM · ~6.7–13 tok/s
Hy3299B MoEOffload Q3_K_M, 128 GB RAM · ~4.0–8.0 tok/s

Comparing cards? The RTX 4070 Ti (launch MSRP $799) is in best GPUs for local LLMs under $800. See also Mac vs PC.

Just too big at 4-bit

These need up to 35% more memory than the RTX 4070 Ti has at Q4_K_M and 8K context. A smaller quant or shorter context may work.

ModelQ4_K_M needsShort by
gpt-oss Safeguard 20B11.8 GB0.4 GB
gpt-oss 20B11.8 GB0.4 GB
DeepSeek MoE 16B Base12.5 GB1.1 GB
ERNIE 4.5 21B A3B PT13.6 GB2.2 GB
ERNIE 4.5 21B A3B Thinking13.6 GB2.2 GB
LFM2.5 24B A2B14.3 GB2.9 GB

Questions

What is the biggest model that runs on the RTX 4070 Ti?

Of the models we track, the largest that fits at 8K context is GLM 4.7 Flash (31B) at Q2_K, needing about 11.4 GB.

How much of the RTX 4070 Ti's memory can a model use?

About 11.4 GB: we assume 95% of the 12 GB is usable after driver and display overhead.

Spec source: NVIDIA specifications. Model data: Hugging Face Hub API. Speed is a bandwidth ceiling (bandwidth ÷ bytes read per token), not a benchmark. Methodology · Report an error.