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

What AI models run on the RX 7900 XT?

The RX 7900 XT has 20 GB of VRAM (800 GB/s). About 19.0 GB is usable for a model. 209 of 298 open models we track fit at 8K context.

209models fit (8K ctx)
197still fit at 32K ctx
49Blargest that fits
800 GB/smemory bandwidth

This card has 20 GB of VRAM. See every model that fits in 24 GB VRAM.

Compare for local LLMs: RX 7900 XT vs RTX 4070 Ti Super all comparisons

Using a specific tool? Ollama models that run on the RX 7900 XT · 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_K747details
Nous Hermes 2 Mixtral 8x7B DPO47B (13B act.)Q2_K17.6 GBno168details
Falcon 40B42BQ2_K16.1 GBQ2_K51details
Flint40B (1.3B act.)Q2_K15.4 GBQ2_K1645details
Maxwell V1 Flash37B (2.3B act.)Q2_K14.0 GBQ2_K940details
Hermes 4.3 36B36BQ2_K15.4 GBno59details
Seed OSS 36B Instruct36BQ2_K15.4 GBno59details
Qwen3.6 35B A3B36B (3.0B act.)Q3_K_M17.0 GBQ2_K560details
Ornith 1.5 35B A3B36B (3.0B act.)Q3_K_M16.6 GBQ3_K_M574details
Qwen3.5 35B A3B36B (3.0B act.)Q3_K_M17.0 GBQ2_K560details
Sage A1.536B (4.7B act.)Q3_K_M17.0 GBQ2_K353details
Qwen AgentWorld 35B A3B35B (3.0B act.)Q3_K_M16.6 GBQ3_K_M554details
Yi 34B 200K34BQ2_K16.0 GBno55details
Yi 34B34BQ2_K16.0 GBno55details
Yi 34B Chat34BQ2_K16.0 GBno55details
Yi 1.5 34B Chat34BQ2_K14.3 GBno62details
Nous Hermes 2 Yi 34B34BQ2_K14.8 GBno60details
DeepSeek R1 Distill Qwen 32B33BQ2_K14.2 GBno65details
Qwen3 32B33BQ2_K14.1 GBno65details
GLM Z1 32B 041433BQ3_K_M16.0 GBQ2_K50details
GLM 4 32B 041433BQ3_K_M16.0 GBQ2_K50details
Olmo 3 32B Think32BQ3_K_M16.4 GBQ2_K51details
Granite 4.0 H Small32B (8.8B act.)Q3_K_M15.0 GBQ3_K_M191details
Gemma 4 31B IT31BQ2_K14.7 GBQ2_K63details
Gemma 4 31B31BQ2_K13.9 GBQ2_K68details
GLM 4.7 Flash31B (4.6B act.)Q3_K_M14.2 GBQ3_K_M385details
Qwen3 Coder 30B A3B Instruct31B (3.0B act.)Q3_K_M15.3 GBQ2_K546details
North Mini Code 1.030B (3.3B act.)Q3_K_M14.6 GBQ3_K_M519details
Hy Memory 30B A3B30B (3.0B act.)Q3_K_M15.1 GBQ2_K546details
Muse Glimmer 30B30BQ3_K_M14.5 GBQ3_K_M55details
Granite 4.2 30B29BQ3_K_M15.8 GBQ2_K57details
Raon 1 Mini Preview29B (4.3B act.)Q3_K_M14.2 GBQ3_K_M379details
Qwen3.8 27B28BQ3_K_M13.9 GBQ3_K_M60details
Qwen3.6 27B28BQ3_K_M15.2 GBQ3_K_M54details
Qwen3.5 27B28BQ3_K_M15.2 GBQ3_K_M54details
Gemma 3 27B IT27BQ3_K_M14.4 GBQ3_K_M60details
Arete 27B27BQ3_K_M13.6 GBQ3_K_M61details
MiniMax H3 Image PE27BQ4_K_M16.8 GBQ3_K_M48details
Trinity Mini26B (2.0B act.)Q4_K_M15.8 GBQ4_K_M667details
GLiDE 27B26BQ4_K_M16.2 GBQ3_K_M50details
DiffusionGemma 26B A4B IT26B (4.0B act.)Q4_K_M17.1 GBQ3_K_M307details
Gemma 4 26B A4B IT26B (4.0B act.)Q3_K_M13.5 GBQ3_K_M396details
LFM2.5 24B A2B24B (2.0B act.)Q5_K_M16.7 GBQ4_K_M563details
ERNIE 4.5 21B A3B PT22BQ5_K_M15.7 GBQ5_K_M51details
ERNIE 4.5 21B A3B Thinking22BQ5_K_M15.7 GBQ5_K_M51details
gpt-oss Safeguard 20B22B (4.8B act.)F1613.8 GBF16260details
gpt-oss 20B21B (4.2B act.)F1613.8 GBF16289details
Kimi VL A3B Thinking 250616B (3.0B act.)Q8_016.8 GBQ6_K258details
Kimi VL A3B Thinking16B (3.0B act.)Q6_K13.5 GBQ6_K325details
Kimi VL A3B Instruct16B (3.0B act.)Q8_016.8 GBQ6_K258details
DeepSeek MoE 16B Base16B (2.8B act.)Q6_K16.0 GBQ3_K_M316details
LLaDA2.1 Mini16B (1.4B act.)Q8_017.1 GBQ6_K525details
Moonlight 16B A3B Instruct16B (3.0B act.)Q8_016.8 GBQ6_K251details
DeepSeek Coder V2 Lite Instruct16B (2.7B act.)Q8_016.4 GBQ8_0283details
DeepSeek V2 Lite16B (2.7B act.)Q8_016.4 GBQ8_0283details
DeepSeek V2 Lite Chat16B (2.7B act.)Q8_016.4 GBQ8_0283details
Qwen2.5 Coder 14B Instruct15BQ8_016.8 GBQ5_K_M51details
DeepSeek R1 Distill Qwen 14B15BQ8_016.8 GBQ5_K_M51details
Qwen3 14B15BQ8_016.6 GBQ6_K51details
Phi 415BQ8_016.7 GBQ5_K_M51details
Gemma 3 12B IT12BQ8_013.4 GBQ8_064details
Llama Guard 4 12B12BQ8_014.2 GBQ6_K63details
KORMo Web 11B11BQ8_013.0 GBQ8_067details
Llama 3.2 11B Vision Instruct11BQ8_011.6 GBQ8_077details
Llama 3.2 11B Vision11BQ8_012.4 GBQ8_071details
GLM 4.6V Flash10BQ8_010.3 GBQ8_080details
GLM 4.1V 9B Thinking10BQ8_010.3 GBQ8_080details
Step3 VL 10B10BF1617.1 GBQ8_049details
SenseNova Loop 9B Preview9.7BQ8_010.0 GBQ8_084details
Qwen3.5 9B9.7BQ8_010.3 GBQ8_082details
Ornith 1.5 9B9.7BQ8_010.0 GBQ8_084details
SenseNova Loop 9B9.4BQ8_010.0 GBQ8_084details
MiMo V2.6 Distill Qwen 9B9.4BQ8_010.0 GBQ8_084details
GLM 4 9B9.4BQ8_010.3 GBQ8_080details
Apertus v1.5 8B8.9BQ8_010.5 GBQ8_085details
InternLM3 8B Instruct8.8BQ8_09.7 GBQ8_085details
Granite 4.2 8B8.8BQ8_010.5 GBQ8_086details
Granite 4.1 8B8.8BQ8_010.5 GBQ8_086details
Granite 4.1 8B Llmon Preview8.8BQ8_010.5 GBQ8_086details
Qwen3 VL 8B Instruct8.8BF1617.1 GBQ8_049details
Molmo2 8B8.7BF1617.1 GBQ8_049details
LFM2.5 8B A1B8.5B (1.0B act.)F1616.5 GBF16400details
Qwen2.5 VL 7B Instruct8.3BF1615.3 GBF1653details
UI TARS 1.5 7B8.3BF1615.3 GBF1653details
olmOCR 2 7B 10258.3BF1615.3 GBF1653details
olmOCR 7B 0225 Preview8.3BF1616.6 GBQ8_048details
UI TARS 7B DPO8.3BF1615.3 GBF1653details
Qwen3 8B8.2BF1617.1 GBQ8_049details
DeepSeek R1 0528 Qwen3 8B8.2BF1617.1 GBQ8_049details
Granite 3.3 8B Instruct8.2BF1617.0 GBQ8_049details
Granite 3.1 8B Instruct8.2BF1617.0 GBQ8_049details
Granite 3.2 8B Instruct8.2BF1617.0 GBQ8_049details
Apertus 8B Instruct 25098.1BF1616.7 GBQ8_050details
Llama 3.1 8B Instruct8.0BF1616.6 GBQ8_050details
Llama 3.1 8B8.0BF1616.6 GBQ8_050details
Hermes 3 Llama 3.1 8B8.0BF1616.6 GBQ8_050details
Llama Guard 3 8B8.0BF1616.6 GBQ8_050details
DeepSeek R1 Distill Llama 8B8.0BF1616.6 GBQ8_050details
Meta Llama 3 8B Instruct8.0BF1616.6 GBQ8_050details
Meta Llama Guard 2 8B8.0BF1616.6 GBQ8_050details
Meta Llama 3.1 8B Instruct8.0BF1616.6 GBQ8_050details
Aya Expanse 8B8.0BF1616.9 GBQ8_050details
Ling 3.0 Tiny7.9B (1.4B act.)F1615.6 GBF16289details
MiMo 7B RL7.8BF1616.4 GBQ8_051details
InternLM2.5 7B Chat7.7BF1616.0 GBQ8_052details
InternLM2 Chat 7B7.7BF1616.0 GBQ8_052details
Qwen2.5 7B Instruct7.6BF1615.3 GBF1653details
DeepSeek R1 Distill Qwen 7B7.6BF1615.3 GBF1653details
Qwen2.5 Coder 7B Instruct7.6BF1615.3 GBF1653details
Olmo 3 7B Instruct7.3BF1616.7 GBQ8_055details
Olmo 3 1025 7B7.3BF1616.7 GBQ8_055details
Olmo 3 7B Think7.3BF1616.7 GBQ8_055details
Mistral 7B v0.17.2BF1614.9 GBQ8_055details
Mistral 7B Instruct v0.27.2BF1614.9 GBQ8_055details
Mistral 7B Instruct v0.17.2BF1614.9 GBQ8_055details
Falcon 7B7.2BF1614.0 GBF1655details
Falcon 7B Instruct7.2BF1614.0 GBF1655details
Granite 4.0 H Tiny6.9B (1.5B act.)F1613.6 GBF16273details
OLMoE 1B 7B 09246.9B (1.3B act.)F1614.4 GBQ8_0312details
DeepSeek Coder 7B Instruct v1.56.9BQ8_011.2 GBQ3_K_M109details
DeepSeek Coder 6.7B Instruct6.7BF1617.0 GBno59details
DeepSeek Coder 6.7B Base6.7BF1617.0 GBno59details
Llama 2 7B HF6.7BF1617.0 GBQ2_K59details
Yi 6B 200K6.1BF1612.3 GBF1666details
Yi 6B6.1BF1612.3 GBF1666details
Molmo2 4B4.9BF1610.0 GBF1691details
WeVisOne VL 4B4.8BF1610.0 GBF1691details
Mage VL Cowork4.7BF1610.0 GBF1691details
Qianfan OCR4.7BF1610.1 GBF1690details
Qwen3.5 4B4.7BF169.2 GBF1692details
SingVL S1 4B4.5BF169.0 GBF1695details
LightOnOCR 3 4B4.5BF169.0 GBF1695details
Intern Decision 4B4.5BF169.0 GBF1695details
Qwen3 VL 4B Instruct4.4BF169.3 GBF1699details
Gemma 3 4B IT4.3BF168.4 GBF16103details
MedGemma 1.5 4B IT4.3BF168.4 GBF16103details
MedGemma 4B IT4.3BF168.4 GBF16103details
TranslateGemma 4B IT4.3BF169.2 GBF1693details
Gemma 3 4B PT4.3BF168.4 GBF16103details
Phi 3.5 Vision Instruct4.1BF1611.2 GBQ8_096details
Qwen3 4B4.0BF169.3 GBF1699details
Qwen3 4B Instruct 25074.0BF169.3 GBF1699details
Granite Vision 4.1 4B4.0BF167.6 GBF16117details
Phi 4 Mini Instruct3.8BF168.9 GBF16104details
Phi 3 Mini 4K Instruct3.8BF1610.6 GBQ8_0105details
Phi 3.5 Mini Instruct3.8BF1610.6 GBQ8_0105details
Phi 3 Mini 128K Instruct3.8BF1610.6 GBQ8_0105details
Qwen2.5 VL 3B Instruct3.8BF168.0 GBF16107details
Granite 4.2 3B3.7BF168.0 GBF16109details
Granite 4.1 3B Llmon Preview3.7BF168.0 GBF16109details
Raon 1 Nano3.5BF167.0 GBF16127details
Granite 4.1 3B3.4BF167.6 GBF16117details
Granite 4.0 Micro3.4BF167.6 GBF16117details
DeepSeek OCR 23.4B (1.2B act.)F166.6 GBF16386details
DeepSeek VL2 Tiny3.4B (1.2B act.)F167.4 GBF16340details
Tiny Aya Global3.3BF167.5 GBF16119details
Unlimited OCR 23.3B (1.1B act.)F166.6 GBF16396details
Unlimited OCR3.3B (1.1B act.)F166.6 GBF16398details
DeepSeek OCR3.3B (1.1B act.)F166.6 GBF16398details
Granite 3.1 3B A800M Instruct3.3B (883M act.)F167.1 GBF16453details
Llama 3.2 3B Instruct3.2BF167.5 GBF16124details
Llama 3.2 3B3.2BF167.5 GBF16125details
Hermes 3 Llama 3.2 3B3.2BF167.5 GBF16124details
d1 3B3.1BF165.8 GBF16148details
LFM2.5 VL 3B3.1BF165.8 GBF16148details
Qwen2.5 3B Instruct3.1BF166.7 GBF16130details
SmolLM3 3B3.1BF166.9 GBF16130details
SmolLM3 3B Base3.1BF166.9 GBF16130details
Phi 22.8BF168.2 GBF16144details
LFM2.5 2.6B2.7BF1610.8 GBF1674details
Granite 3.3 2B Instruct2.5BF165.8 GBF16158details
MiniCPM5 2B2.5BF165.7 GBF16159details
WeVisOne VL 2B2.4BF165.4 GBF16196details
Cosmos Reason2 2B2.4BF165.4 GBF16197details
Qwen3.5 2B2.3BF164.6 GBF16205details
SmolVLM Instruct2.2BF165.4 GBF16221details
Qwen3 1.7B2.0BF165.4 GBF16197details
d1 2B1.9BF164.0 GBF16231details
LFM3 VL 2B1.9BF164.0 GBF16231details
DeepSeek R1 Distill Qwen 1.5B1.8BF164.2 GBF16225details
LFM3 1.7B1.7BF167.2 GBF16115details
LFM3 1.7B Base1.7BF164.0 GBF16231details
SmolLM2 1.7B Instruct1.7BF165.2 GBF16234details
SmolLM2 1.7B1.7BF165.2 GBF16234details
SmolLM 1.7B1.7BF165.2 GBF16234details
Granite 4.0 1B1.6BF164.3 GBF16245details
Qwen2.5 1.5B Instruct1.5BF163.8 GBF16259details
OLMo 2 0425 1B1.5BF164.4 GBF16269details
DeepSeek Coder 1.3B Instruct1.3BF164.5 GBF16297details
Llama 3.2 1B1.2BF163.2 GBF16323details
Llama 3.2 1B Instruct1.2BF163.2 GBF16323details
LFM2.5 1.2B Instruct1.2BF165.0 GBF16171details
LFM2 1.2B1.2BF163.0 GBF16341details
HunyuanOCR1.1BF163.5 GBF16357details
PaddleOCR VL 1.6959MF162.5 GBF16417details
Qwen3.5 0.8B873MF162.4 GBF16514details
LightOnOCR 3 0.8B853MF162.4 GBF16527details
FastVLM 0.5B759MF162.2 GBF16527details
Qwen3 0.6B752MF163.0 GBF16530details
GOT OCR2.0716MF162.8 GBF16559details
Nemotron 3.5 Parse631MF161.9 GBF16612details
GOT OCR 2.0 HF561MF162.5 GBF16714details
SmolVLM2 500M Video Instruct507MF161.6 GBF16975details
SmolVLM 500M Instruct507MF161.6 GBF16975details
Qwen2.5 0.5B Instruct494MF161.7 GBF16810details
LFM2.5 VL 450M449MF161.3 GBF161124details
SmolLM2 360M Instruct362MF161.5 GBF161103details
SmolLM2 360M362MF161.5 GBF161103details
LFM2.5 350M354MF161.9 GBF16562details

Bigger models with CPU offload

These popular models don't fit in the RX 7900 XT'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 · ~7.6–15 tok/s
gpt-oss 120B117B MoEOffload Q4_K_M, 64 GB RAM · ~16–31 tok/s
DeepSeek V4 Flash 0731304B MoEOffload Q2_K, 128 GB RAM · ~8.6–16 tok/s
DeepSeek V4 Flash291B MoEOffload Q3_K_M, 128 GB RAM · ~7.4–15 tok/s
GLM 4.5 Air110B MoEOffload Q4_K_M, 64 GB RAM · ~6.1–12 tok/s
MiniMax M2.5229B MoEOffload Q4_K_M, 128 GB RAM · ~5.8–12 tok/s
Hy3299B MoEOffload Q3_K_M, 128 GB RAM · ~4.4–8.8 tok/s

Comparing cards? The RX 7900 XT (launch MSRP $899) is in best GPUs for local LLMs under $1,000. See also Mac vs PC.

Just too big at 4-bit

These need up to 35% more memory than the RX 7900 XT has at Q4_K_M and 8K context. A smaller quant or shorter context may work.

ModelQ4_K_M needsShort by
Granite 4.2 30B19.1 GB0.1 GB
Raon 1 Mini Preview19.4 GB0.4 GB
GLM 4 32B 041419.5 GB0.5 GB
GLM Z1 32B 041419.5 GB0.5 GB
Olmo 3 32B Think20.0 GB1.0 GB
Gemma 4 31B20.7 GB1.7 GB
Qwen3 32B21.1 GB2.1 GB
DeepSeek R1 Distill Qwen 32B21.2 GB2.2 GB
Gemma 4 31B IT21.2 GB2.2 GB
Sage A1.521.3 GB2.3 GB
Ornith 1.5 35B A3B21.4 GB2.4 GB
Yi 34B21.6 GB2.6 GB
Yi 34B 200K21.6 GB2.6 GB
Yi 34B Chat21.6 GB2.6 GB
Yi 1.5 34B Chat21.6 GB2.6 GB

Questions

What is the biggest model that runs on the RX 7900 XT?

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 RX 7900 XT's memory can a model use?

About 19.0 GB: we assume 95% of the 20 GB is usable after driver and display overhead.

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