RTX 4090 — what LLMs can it run?
84 of 110 indexed models fit comfortably in 24GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for RTX 4090
24 GB, 1,008 GB/s GDDR6X. 84 of 110 models in this index fit comfortably at 4K context; 87 load at all.
- Biggest that fits
- Seed-OSS 36B Instruct — 19.9 GB at AWQ INT4 — 4.1 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- ERNIE 4.5 21B-A3B — 14.1 GB at Q4_K_M — under 60% of the card, which leaves 10.0 GB for a long context window or a second process.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 1,008 GB/s puts a hard ceiling near 218 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. This index has 13 measured runs on this card; see the table below for what they actually returned.
- The wall you will hit
- Capacity. 84 of 110 models clear 24 GB comfortably at 4K context, and the ones that do not are short by whole gigabytes rather than by a rounding error — at 1,008 GB/s the memory system is not what holds this card back. Moving up a size class is what changes the list, not a different quant.
32B · 27
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Seed-OSS 36B Instruct36B | AWQ INT4 | 19.91 GB | +4.1 GB | ≤59Ceiling |
| Command R 35B35B | GPTQ INT4 | 18.97 GB | +5 GB | ≤61Ceiling |
| Yi 1.5 34B Chat34B | AWQ INT4 | 19 GB | +5 GB | ≤62Ceiling |
| Qwen3 32B Instruct32B | AWQ INT4 | 18.22 GB | +5.8 GB | ≤65Ceiling |
| Qwen2.5 32B Instruct32B | EXL2 3.5bpw | 15.96 GB | +8 GB | ≤75Ceiling |
| Qwen2.5-Coder 32B Instruct32B | AWQ INT4 | 18.08 GB | +5.9 GB | ≤65Ceiling |
| DeepSeek-R1-Distill-Qwen-32B32B | EXL2 3.5bpw | 15.96 GB | +8 GB | ≤75Ceiling |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q4_0 | 19.79 GB | +4.2 GB | ≤591Ceiling |
| Gemma 4 31B IT31B | Q4_K_M | 21.08 GB | +2.9 GB | 38–50Derived |
| Qwen3 30B-A3B Instruct30B-A3B | Q4_K_M | 19.73 GB | +4.3 GB | 95Measured |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q4_K_M | 19.73 GB | +4.3 GB | 92Measured |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q4_K_M | 19.73 GB | +4.3 GB | ≤583Ceiling |
| GLM-4.7-Flash30B-A3B | Q4_K_M | 19.19 GB | +4.8 GB | ≤584Ceiling |
| Granite 4.2 30B30B | Q4_K_M | 19.61 GB | +4.4 GB | 40–53Derived |
| Muse Glimmer 30B30B | Q4_K_M | 17.66 GB | +6.3 GB | 42–56Derived |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +6.5 GB | 43–57Derived |
| Qwen3.6 27B27B | Q4_K_M | 19.07 GB | +4.9 GB | 39–52Derived |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +5.6 GB | 42–56Derived |
| Gemma 2 27B Instruct27B | Q4_K_M | 18.81 GB | +5.2 GB | 43–57Derived |
| Gemma 4 26B-A4B IT26B-A4B | Q4_K_M | 16.39 GB | +7.6 GB | ≤438Ceiling |
| Mistral Small 24B Instruct24B | EXL2 4.65bpw | 15.26 GB | +8.7 GB | ≤76Ceiling |
| LFM2.5 24B-A2B24B-A2B | Q4_K_M | 15.27 GB | +8.7 GB | ≤715Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +3.1 GB | 36–48Derived |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +3.1 GB | 36–48Derived |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +9 GB | 52–69Derived |
| ERNIE 4.5 21B-A3B21B-A3B | Q4_K_M | 14.05 GB | +10 GB | ≤584Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +11.2 GB | ≤509Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +8.4 GB | 50–66Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +6.4 GB | ≤416Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +13.9 GB | ≤729Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +13.8 GB | 78–103Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +12.4 GB | 67–89Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +12.3 GB | 68–89Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | EXL2 4.65bpw | 9.75 GB | +14.3 GB | ≤124Ceiling |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +12.2 GB | 68–89Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +11.1 GB | 61–81Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +12.9 GB | 71–93Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +14.2 GB | 81–108Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +10.4 GB | ≤400Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +15.7 GB | 96–127Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +11.1 GB | 60–80Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +16.4 GB | 109–144Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +16.7 GB | 112–148Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +12.3 GB | 72–95Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +13.6 GB | 72–95Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +13.8 GB | 74–98Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +13.6 GB | 75–100Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +13.6 GB | 76–100Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +14 GB | 78–103Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +18.5 GB | 140–186Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +18.1 GB | 142–188Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +16.4 GB | 105–139Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +14.5 GB | 83–109Derived |
| Nous Hermes 3 Llama 3.1 8B8B | EXL2 4.65bpw | 5.43 GB | +18.6 GB | ≤228Ceiling |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +18.4 GB | 145–192Derived |
| OpenChat 3.6 8B8B | EXL2 4.65bpw | 5.43 GB | +18.6 GB | ≤228Ceiling |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +17.5 GB | 124–164Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +18.6 GB | 150–199Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +17.2 GB | 113–149Derived |
| Qwen2.5-Coder 7B Instruct7B | EXL2 4.65bpw | 4.87 GB | +19.1 GB | ≤239Ceiling |
| DeepSeek-R1-Distill-Qwen-7B7B | EXL2 4.65bpw | 4.87 GB | +19.1 GB | ≤239Ceiling |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +15.5 GB | 88–117Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +13.7 GB | 91–120Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +13.7 GB | 91–120Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +18.9 GB | 161–213Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +17.3 GB | 119–157Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +17.3 GB | 119–157Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +18.8 GB | ≤215Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +19 GB | ≤230Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +19 GB | ≤232Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +20 GB | ≤323Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +19.2 GB | ≤260Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +18.1 GB | ≤261Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +19.6 GB | ≤273Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +20 GB | ≤311Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +20.4 GB | ≤323Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +20.3 GB | ≤324Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +20.7 GB | ≤383Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +21.8 GB | ≤509Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +21.6 GB | ≤579Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +22.2 GB | ≤650Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +22.5 GB | ≤806Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +23.1 GB | ≤1244Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +23.4 GB | ≤2057Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
A further 3 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.
Measured on this card
13 benchmark runs in this index were recorded on a RTX 4090. Everything else on this page is calculated, not observed.
| Model | Stack | tok/s | VRAM used |
|---|---|---|---|
| Llama 3.1 8B | ExLlamaV2 · EXL2 4.65bpw | 235 | 5.4 GB |
| Llama 3.1 8B | vLLM · AWQ INT4 | 218 | 4.9 GB |
| Llama 3.1 8B | llama.cpp · GGUF Q4_K_M | 148 | 5.7 GB |
| Qwen2.5 7B | llama.cpp · GGUF Q4_K_M | 155 | 5.4 GB |
| DeepSeek-R1 14B | ExLlamaV2 · EXL2 4.65bpw | 128 | 9.8 GB |
| Qwen2.5 32B | llama.cpp · GGUF Q4_K_M | 44 | 22 GB |
| Qwen3 8B | llama.cpp · GGUF Q4_K_M | 142 | 5.8 GB |
| Qwen3 14B | ExLlamaV2 · EXL2 4.65bpw | 118 | 10 GB |
| Qwen3 32B | llama.cpp · GGUF Q4_K_M | 42 | 22.5 GB |
| Qwen3 30B-A3B | llama.cpp · GGUF Q4_K_M | 95 | 19 GB |
| R1-Distill-Llama-8B | llama.cpp · GGUF Q4_K_M | 145 | 5.6 GB |
| Phi-4 14B | llama.cpp · GGUF Q4_K_M | 88 | 9.1 GB |
| Qwen3-Coder 30B-A3B | llama.cpp · GGUF Q4_K_M | 92 | 19.2 GB |
Cards with the same budget
What fits is decided by memory, so every 24GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.
Same list, different speed — play one sentence on both cards: RTX 4090 vs RTX 3090 (Gemma 4 31B IT)
Setup guides for this card
Step-by-step guides written for the RTX 4090's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A RTX 5090 (32GB) fits 4 more of the indexed models than this card. RTX 5090 →
Before buying, hear the speed difference on a model both cards hold: RTX 4090 vs RTX 5090 (Gemma 4 31B IT) →
Common questions
What is the best local LLM for an RTX 4090?
For everyday use, ERNIE 4.5 21B-A3B at Q4_K_M — about 14.1 GB of the card's 24 GB at 4K context, leaving 10.0 GB for a longer window. If you want the largest thing that will load, that is Seed-OSS 36B Instruct at AWQ INT4 (19.9 GB, 4.1 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can an RTX 4090 run Mixtral 8x7B Instruct?
Only just. Its smallest build here, AWQ INT4, needs about 24.9 GB against 24.0 GB usable — that loads with nothing else running, with no margin for a longer context window. It is not on the list above, which requires a model to stay inside 88% of usable capacity.
How many tokens per second does an RTX 4090 do on an 8B model at Q4?
This index has 13 measured runs on this card — those figures are in the table on this page. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 1,008 GB/s — a ceiling near 218 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
RTX 4090 or RTX 3090 for local LLMs?
They hold the same models: 24 GB against 24 GB fits 84 and 84 of 110 respectively at 4K. The difference is throughput — 1,008 GB/s against 936 GB/s, a 1.1× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 4090 generates faster on any model both can hold.
84 of 110 indexed models fit comfortably in 24GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml