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

What AI models run on the RTX 4090?

The RTX 4090 has 24 GB of VRAM (1,008 GB/s). About 22.8 GB is usable for a model. 209 of 298 open models we track fit at 8K context.

209models fit (8K ctx)
209still fit at 32K ctx
49Blargest that fits
1008 GB/smemory bandwidth

This card has 24 GB of VRAM. See every model that fits in 24 GB VRAM. Other 24 GB hardware: RTX 3090, RTX 3090 Ti, RX 7900 XTX, Mac M2 24GB, Mac M3 24GB, Mac M4 Pro 24GB.

Compare for local LLMs: RTX 4090 vs RTX 4080RTX 4090 vs RTX 3090RTX 4090 vs RTX 5090RTX 4090 vs RX 7900 XTXRTX 4090 vs Mac M4 Pro 48GB all comparisons

Using a specific tool? Ollama models that run on the RTX 4090 · LM Studio · llama.cpp offload

Models that fit, largest first

ModelParamsBest quant (8K)Needs32K contextMax tok/s (est.)Details
Kimi Linear 48B A3B Instruct49B (3.0B act.)Q2_K17.3 GBQ2_K941details
Nous Hermes 2 Mixtral 8x7B DPO47B (13B act.)Q2_K17.6 GBQ2_K211details
Falcon 40B42BQ3_K_M20.2 GBQ2_K50details
Flint40B (1.3B act.)Q3_K_M19.1 GBQ3_K_M1643details
Maxwell V1 Flash37B (2.3B act.)Q3_K_M17.9 GBQ3_K_M908details
Hermes 4.3 36B36BQ3_K_M19.1 GBQ2_K57details
Seed OSS 36B Instruct36BQ3_K_M19.1 GBQ2_K57details
Qwen3.6 35B A3B36B (3.0B act.)Q3_K_M17.0 GBQ3_K_M706details
Ornith 1.5 35B A3B36B (3.0B act.)Q3_K_M16.6 GBQ3_K_M723details
Qwen3.5 35B A3B36B (3.0B act.)Q3_K_M17.0 GBQ3_K_M706details
Sage A1.536B (4.7B act.)Q3_K_M17.0 GBQ3_K_M445details
Qwen AgentWorld 35B A3B35B (3.0B act.)Q3_K_M16.6 GBQ3_K_M698details
Yi 34B 200K34BQ3_K_M17.9 GBQ2_K61details
Yi 34B34BQ3_K_M17.9 GBQ2_K61details
Yi 34B Chat34BQ3_K_M17.9 GBQ2_K61details
Yi 1.5 34B Chat34BQ3_K_M18.1 GBQ2_K60details
Nous Hermes 2 Yi 34B34BQ3_K_M18.4 GBQ2_K59details
DeepSeek R1 Distill Qwen 32B33BQ3_K_M17.5 GBQ2_K63details
Qwen3 32B33BQ3_K_M17.6 GBQ2_K63details
GLM Z1 32B 041433BQ4_K_M19.5 GBQ3_K_M51details
GLM 4 32B 041433BQ4_K_M19.5 GBQ3_K_M51details
Olmo 3 32B Think32BQ4_K_M20.0 GBQ3_K_M52details
Granite 4.0 H Small32B (8.8B act.)Q4_K_M18.9 GBQ4_K_M189details
Gemma 4 31B IT31BQ3_K_M17.8 GBQ2_K63details
Gemma 4 31B31BQ3_K_M17.2 GBQ2_K66details
GLM 4.7 Flash31B (4.6B act.)Q4_K_M18.3 GBQ4_K_M370details
Qwen3 Coder 30B A3B Instruct31B (3.0B act.)Q4_K_M18.8 GBQ3_K_M551details
North Mini Code 1.030B (3.3B act.)Q4_K_M18.8 GBQ4_K_M496details
Hy Memory 30B A3B30B (3.0B act.)Q4_K_M18.5 GBQ3_K_M550details
Muse Glimmer 30B30BQ4_K_M17.9 GBQ4_K_M55details
Granite 4.2 30B29BQ4_K_M19.1 GBQ2_K57details
Raon 1 Mini Preview29B (4.3B act.)Q4_K_M19.4 GBQ4_K_M341details
Qwen3.8 27B28BQ4_K_M17.6 GBQ4_K_M58details
Qwen3.6 27B28BQ4_K_M18.1 GBQ4_K_M56details
Qwen3.5 27B28BQ4_K_M18.1 GBQ4_K_M56details
Gemma 3 27B IT27BQ5_K_M19.9 GBQ4_K_M52details
Arete 27B27BQ4_K_M17.4 GBQ4_K_M59details
MiniMax H3 Image PE27BQ5_K_M19.3 GBQ4_K_M52details
Trinity Mini26B (2.0B act.)Q5_K_M18.4 GBQ5_K_M719details
GLiDE 27B26BQ5_K_M18.7 GBQ5_K_M54details
DiffusionGemma 26B A4B IT26B (4.0B act.)Q5_K_M19.2 GBQ5_K_M340details
Gemma 4 26B A4B IT26B (4.0B act.)Q5_K_M19.4 GBQ5_K_M337details
LFM2.5 24B A2B24B (2.0B act.)Q6_K19.1 GBQ6_K614details
ERNIE 4.5 21B A3B PT22BQ6_K17.9 GBQ6_K56details
ERNIE 4.5 21B A3B Thinking22BQ6_K17.9 GBQ6_K56details
gpt-oss Safeguard 20B22B (4.8B act.)F1613.8 GBF16328details
gpt-oss 20B21B (4.2B act.)F1613.8 GBF16365details
Kimi VL A3B Thinking 250616B (3.0B act.)Q8_016.8 GBQ8_0325details
Kimi VL A3B Thinking16B (3.0B act.)Q8_017.2 GBQ8_0316details
Kimi VL A3B Instruct16B (3.0B act.)Q8_016.8 GBQ8_0325details
DeepSeek MoE 16B Base16B (2.8B act.)Q8_018.6 GBQ5_K_M335details
LLaDA2.1 Mini16B (1.4B act.)Q8_017.1 GBQ8_0662details
Moonlight 16B A3B Instruct16B (3.0B act.)Q8_016.8 GBQ8_0316details
DeepSeek Coder V2 Lite Instruct16B (2.7B act.)Q8_016.4 GBQ8_0356details
DeepSeek V2 Lite16B (2.7B act.)Q8_016.4 GBQ8_0357details
DeepSeek V2 Lite Chat16B (2.7B act.)Q8_016.4 GBQ8_0356details
Qwen2.5 Coder 14B Instruct15BQ8_016.8 GBQ6_K64details
DeepSeek R1 Distill Qwen 14B15BQ8_016.8 GBQ6_K64details
Qwen3 14B15BQ8_016.6 GBQ8_064details
Phi 415BQ8_016.7 GBQ6_K65details
Gemma 3 12B IT12BQ8_013.4 GBQ8_081details
Llama Guard 4 12B12BQ8_014.2 GBQ8_079details
KORMo Web 11B11BQ8_013.0 GBQ8_085details
Llama 3.2 11B Vision Instruct11BF1620.1 GBQ8_052details
Llama 3.2 11B Vision11BQ8_012.4 GBQ8_089details
GLM 4.6V Flash10BF1618.5 GBF1654details
GLM 4.1V 9B Thinking10BF1618.5 GBF1654details
Step3 VL 10B10BF1617.1 GBF1662details
SenseNova Loop 9B Preview9.7BF1617.8 GBF1656details
Qwen3.5 9B9.7BF1618.3 GBF1655details
Ornith 1.5 9B9.7BF1617.8 GBF1656details
SenseNova Loop 9B9.4BF1618.7 GBF1654details
MiMo V2.6 Distill Qwen 9B9.4BF1617.8 GBF1656details
GLM 4 9B9.4BF1618.5 GBF1654details
Apertus v1.5 8B8.9BF1618.2 GBQ8_057details
InternLM3 8B Instruct8.8BF1617.4 GBF1657details
Granite 4.2 8B8.8BF1618.2 GBQ8_057details
Granite 4.1 8B8.8BF1618.2 GBQ8_057details
Granite 4.1 8B Llmon Preview8.8BF1618.2 GBQ8_057details
Qwen3 VL 8B Instruct8.8BF1617.1 GBF1662details
Molmo2 8B8.7BF1617.1 GBF1662details
LFM2.5 8B A1B8.5B (1.0B act.)F1616.5 GBF16504details
Qwen2.5 VL 7B Instruct8.3BF1615.3 GBF1666details
UI TARS 1.5 7B8.3BF1615.3 GBF1666details
olmOCR 2 7B 10258.3BF1615.3 GBF1666details
olmOCR 7B 0225 Preview8.3BF1616.6 GBF1661details
UI TARS 7B DPO8.3BF1615.3 GBF1666details
Qwen3 8B8.2BF1617.1 GBF1662details
DeepSeek R1 0528 Qwen3 8B8.2BF1617.1 GBF1662details
Granite 3.3 8B Instruct8.2BF1617.0 GBQ8_062details
Granite 3.1 8B Instruct8.2BF1617.0 GBQ8_062details
Granite 3.2 8B Instruct8.2BF1617.0 GBQ8_062details
Apertus 8B Instruct 25098.1BF1616.7 GBF1663details
Llama 3.1 8B Instruct8.0BF1616.6 GBF1663details
Llama 3.1 8B8.0BF1616.6 GBF1663details
Hermes 3 Llama 3.1 8B8.0BF1616.6 GBF1663details
Llama Guard 3 8B8.0BF1616.6 GBF1663details
DeepSeek R1 Distill Llama 8B8.0BF1616.6 GBF1663details
Meta Llama 3 8B Instruct8.0BF1616.6 GBF1663details
Meta Llama Guard 2 8B8.0BF1616.6 GBF1663details
Meta Llama 3.1 8B Instruct8.0BF1616.6 GBF1663details
Aya Expanse 8B8.0BF1616.9 GBF1663details
Ling 3.0 Tiny7.9B (1.4B act.)F1615.6 GBF16364details
MiMo 7B RL7.8BF1616.4 GBF1664details
InternLM2.5 7B Chat7.7BF1616.0 GBF1665details
InternLM2 Chat 7B7.7BF1616.0 GBF1665details
Qwen2.5 7B Instruct7.6BF1615.3 GBF1666details
DeepSeek R1 Distill Qwen 7B7.6BF1615.3 GBF1666details
Qwen2.5 Coder 7B Instruct7.6BF1615.3 GBF1666details
Olmo 3 7B Instruct7.3BF1616.7 GBF1669details
Olmo 3 1025 7B7.3BF1616.7 GBF1669details
Olmo 3 7B Think7.3BF1616.7 GBF1669details
Mistral 7B v0.17.2BF1614.9 GBF1670details
Mistral 7B Instruct v0.27.2BF1614.9 GBF1670details
Mistral 7B Instruct v0.17.2BF1614.9 GBF1670details
Falcon 7B7.2BF1614.0 GBF1670details
Falcon 7B Instruct7.2BF1614.0 GBF1670details
Granite 4.0 H Tiny6.9B (1.5B act.)F1613.6 GBF16344details
OLMoE 1B 7B 09246.9B (1.3B act.)F1614.4 GBF16393details
DeepSeek Coder 7B Instruct v1.56.9BF1617.2 GBQ5_K_M73details
DeepSeek Coder 6.7B Instruct6.7BF1617.0 GBQ4_K_M75details
DeepSeek Coder 6.7B Base6.7BF1617.0 GBQ4_K_M75details
Llama 2 7B HF6.7BF1617.0 GBQ4_K_M75details
Yi 6B 200K6.1BF1612.3 GBF1683details
Yi 6B6.1BF1612.3 GBF1683details
Molmo2 4B4.9BF1610.0 GBF16114details
WeVisOne VL 4B4.8BF1610.0 GBF16114details
Mage VL Cowork4.7BF1610.0 GBF16114details
Qianfan OCR4.7BF1610.1 GBF16114details
Qwen3.5 4B4.7BF169.2 GBF16116details
SingVL S1 4B4.5BF169.0 GBF16120details
LightOnOCR 3 4B4.5BF169.0 GBF16120details
Intern Decision 4B4.5BF169.0 GBF16120details
Qwen3 VL 4B Instruct4.4BF169.3 GBF16125details
Gemma 3 4B IT4.3BF168.4 GBF16130details
MedGemma 1.5 4B IT4.3BF168.4 GBF16130details
MedGemma 4B IT4.3BF168.4 GBF16130details
TranslateGemma 4B IT4.3BF169.2 GBF16117details
Gemma 3 4B PT4.3BF168.4 GBF16130details
Phi 3.5 Vision Instruct4.1BF1611.2 GBF16122details
Qwen3 4B4.0BF169.3 GBF16125details
Qwen3 4B Instruct 25074.0BF169.3 GBF16125details
Granite Vision 4.1 4B4.0BF167.6 GBF16148details
Phi 4 Mini Instruct3.8BF168.9 GBF16131details
Phi 3 Mini 4K Instruct3.8BF1610.6 GBF16132details
Phi 3.5 Mini Instruct3.8BF1610.6 GBF16132details
Phi 3 Mini 128K Instruct3.8BF1610.6 GBF16132details
Qwen2.5 VL 3B Instruct3.8BF168.0 GBF16134details
Granite 4.2 3B3.7BF168.0 GBF16138details
Granite 4.1 3B Llmon Preview3.7BF168.0 GBF16138details
Raon 1 Nano3.5BF167.0 GBF16161details
Granite 4.1 3B3.4BF167.6 GBF16148details
Granite 4.0 Micro3.4BF167.6 GBF16148details
DeepSeek OCR 23.4B (1.2B act.)F166.6 GBF16487details
DeepSeek VL2 Tiny3.4B (1.2B act.)F167.4 GBF16429details
Tiny Aya Global3.3BF167.5 GBF16150details
Unlimited OCR 23.3B (1.1B act.)F166.6 GBF16499details
Unlimited OCR3.3B (1.1B act.)F166.6 GBF16502details
DeepSeek OCR3.3B (1.1B act.)F166.6 GBF16502details
Granite 3.1 3B A800M Instruct3.3B (883M act.)F167.1 GBF16570details
Llama 3.2 3B Instruct3.2BF167.5 GBF16157details
Llama 3.2 3B3.2BF167.5 GBF16157details
Hermes 3 Llama 3.2 3B3.2BF167.5 GBF16157details
d1 3B3.1BF165.8 GBF16187details
LFM2.5 VL 3B3.1BF165.8 GBF16187details
Qwen2.5 3B Instruct3.1BF166.7 GBF16163details
SmolLM3 3B3.1BF166.9 GBF16164details
SmolLM3 3B Base3.1BF166.9 GBF16164details
Phi 22.8BF168.2 GBF16181details
LFM2.5 2.6B2.7BF1610.8 GBF1693details
Granite 3.3 2B Instruct2.5BF165.8 GBF16199details
MiniCPM5 2B2.5BF165.7 GBF16200details
WeVisOne VL 2B2.4BF165.4 GBF16247details
Cosmos Reason2 2B2.4BF165.4 GBF16248details
Qwen3.5 2B2.3BF164.6 GBF16259details
SmolVLM Instruct2.2BF165.4 GBF16278details
Qwen3 1.7B2.0BF165.4 GBF16248details
d1 2B1.9BF164.0 GBF16291details
LFM3 VL 2B1.9BF164.0 GBF16291details
DeepSeek R1 Distill Qwen 1.5B1.8BF164.2 GBF16283details
LFM3 1.7B1.7BF167.2 GBF16145details
LFM3 1.7B Base1.7BF164.0 GBF16291details
SmolLM2 1.7B Instruct1.7BF165.2 GBF16294details
SmolLM2 1.7B1.7BF165.2 GBF16294details
SmolLM 1.7B1.7BF165.2 GBF16294details
Granite 4.0 1B1.6BF164.3 GBF16309details
Qwen2.5 1.5B Instruct1.5BF163.8 GBF16326details
OLMo 2 0425 1B1.5BF164.4 GBF16339details
DeepSeek Coder 1.3B Instruct1.3BF164.5 GBF16374details
Llama 3.2 1B1.2BF163.2 GBF16407details
Llama 3.2 1B Instruct1.2BF163.2 GBF16407details
LFM2.5 1.2B Instruct1.2BF165.0 GBF16215details
LFM2 1.2B1.2BF163.0 GBF16430details
HunyuanOCR1.1BF163.5 GBF16450details
PaddleOCR VL 1.6959MF162.5 GBF16526details
Qwen3.5 0.8B873MF162.4 GBF16647details
LightOnOCR 3 0.8B853MF162.4 GBF16665details
FastVLM 0.5B759MF162.2 GBF16664details
Qwen3 0.6B752MF163.0 GBF16668details
GOT OCR2.0716MF162.8 GBF16704details
Nemotron 3.5 Parse631MF161.9 GBF16771details
GOT OCR 2.0 HF561MF162.5 GBF16899details
SmolVLM2 500M Video Instruct507MF161.6 GBF161229details
SmolVLM 500M Instruct507MF161.6 GBF161229details
Qwen2.5 0.5B Instruct494MF161.7 GBF161020details
LFM2.5 VL 450M449MF161.3 GBF161417details
SmolLM2 360M Instruct362MF161.5 GBF161389details
SmolLM2 360M362MF161.5 GBF161389details
LFM2.5 350M354MF161.9 GBF16708details

Bigger models with CPU offload

These popular models don't fit in the RTX 4090'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 · ~8.1–16 tok/s
gpt-oss 120B117B MoEOffload Q4_K_M, 64 GB RAM · ~17–34 tok/s
DeepSeek V4 Flash 0731304B MoEOffload Q3_K_M, 128 GB RAM · ~6.8–13 tok/s
DeepSeek V4 Flash291B MoEOffload Q3_K_M, 128 GB RAM · ~7.7–15 tok/s
GLM 4.5 Air110B MoEOffload Q4_K_M, 64 GB RAM · ~6.5–13 tok/s
MiniMax M2.5229B MoEOffload Q4_K_M, 128 GB RAM · ~6.1–12 tok/s
Hy3299B MoEOffload Q3_K_M, 128 GB RAM · ~4.5–9.1 tok/s

Comparing cards? The RTX 4090 (launch MSRP $1599) is in best GPUs for local LLMs under $2,000. See also Mac vs PC.

Just too big at 4-bit

These need up to 35% more memory than the RTX 4090 has at Q4_K_M and 8K context. A smaller quant or shorter context may work.

ModelQ4_K_M needsShort by
Hermes 4.3 36B23.0 GB0.2 GB
Seed OSS 36B Instruct23.0 GB0.2 GB
Flint24.5 GB1.7 GB
Falcon 40B25.2 GB2.4 GB
Nous Hermes 2 Mixtral 8x7B DPO28.0 GB5.2 GB
Kimi Linear 48B A3B Instruct28.9 GB6.1 GB

Questions

What is the biggest model that runs on the RTX 4090?

Of the models we track, the largest that fits at 8K context is Kimi Linear 48B A3B Instruct (49B) at Q2_K, needing about 17.3 GB.

How much of the RTX 4090's memory can a model use?

About 22.8 GB: we assume 95% of the 24 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.