LLM Roofline
NVIDIA · PCIE · Backend: vllm

NVIDIA RTX 4090 (24 GB) — LLM Sizing & Speed Matrix

1,008 GB/s 24 GB workhorse; fits 32B Q4_K_M tight or 24B Q5_K_M with fast prefill.

Reference Pricing
$1,999 Street
Usable Memory
23 GiB
Memory Bandwidth
1008 GB/s (82% eff.)
FP16 Compute
165.2 TFLOPS
Roofline Ridge
129.9 FLOP/B
Power (Load / Idle)
450W / 35W

Buy or Rent NVIDIA RTX 4090 (24 GB)

Vendor & Cloud Links
Deploy on Vast.ai Spot GPU
Vast.ai · On-Demand / Spot GPU
↗
Launch Pod on RunPod
RunPod · On-Demand / Spot GPU
↗
Run Serverless / GPU on Novita AI
Novita AI · On-Demand / Spot GPU
↗
RTX 4090 24 GB on Amazon
Amazon · Direct Hardware
↗

All 32 Open-Weight Models on NVIDIA RTX 4090 (24 GB) (Q4_K_M @ 32k)

Model Status Weights Total @ 32k Decode TTFT (4k) Max Safe Ctx
Gemma 3 27B IT COMFORTABLE 15.8 GiB 20.2 GiB 41.2 tok/s 2.21s 68k
Mistral Small 3.1 24B COMFORTABLE 13.8 GiB 20.3 GiB 40.9 tok/s 1.92s 49k
Qwen3 14B COMFORTABLE 8.5 GiB 15.0 GiB 56.9 tok/s 1.23s 40k
Qwen2.5 14B Instruct COMFORTABLE 8.5 GiB 16.0 GiB 53.0 tok/s 1.24s 32k
DeepSeek-R1 Distill Qwen 14B COMFORTABLE 8.5 GiB 16.0 GiB 53.0 tok/s 1.24s 69k
Phi-4 14B COMFORTABLE 8.5 GiB 13.0 GiB 66.4 tok/s 1.22s 16k
Gemma 3 12B IT COMFORTABLE 7.0 GiB 10.6 GiB 82.4 tok/s 1.03s 128k
Qwen3 8B COMFORTABLE 4.7 GiB 10.5 GiB 83.5 tok/s 0.71s 40k
Llama 3.1 8B Instruct COMFORTABLE 4.6 GiB 9.9 GiB 89.3 tok/s 0.69s 128k
Qwen2.5 7B Instruct COMFORTABLE 4.4 GiB 7.4 GiB 125.6 tok/s 0.65s 32k
Gemma 3 4B IT COMFORTABLE 2.5 GiB 4.4 GiB 230.8 tok/s 0.39s 128k
GLM-4 32B 0414 TIGHT 18.6 GiB 22.1 GiB 37.5 tok/s 2.62s 32k
Qwen3 30B-A3B TIGHT 17.6 GiB 21.6 GiB 138.2 tok/s 0.35s 40k
Codestral 22B TIGHT 12.8 GiB 21.4 GiB 38.9 tok/s 1.84s 32k
DeepSeek-R1 671B OOM 386.4 GiB 390.3 GiB OOM — OOM
DeepSeek-V3 671B OOM 386.4 GiB 390.3 GiB OOM — OOM
Llama 3.1 405B Instruct OOM 233.8 GiB 252.5 GiB OOM — OOM
Llama 4 Maverick 400B-A17B OOM 231.5 GiB 239.0 GiB OOM — OOM
Qwen3 235B-A22B OOM 135.4 GiB 142.6 GiB OOM — OOM
Mixtral 8x22B Instruct OOM 81.2 GiB 89.8 GiB OOM — OOM
Command A 111B OOM 63.9 GiB 69.1 GiB OOM — OOM
Llama 4 Scout 109B-A17B OOM 62.8 GiB 70.2 GiB OOM — OOM
Qwen2.5 72B Instruct OOM 41.9 GiB 53.7 GiB OOM — OOM
DeepSeek-R1 Distill Llama 70B OOM 40.7 GiB 52.5 GiB OOM — OOM
Llama 3.3 70B Instruct OOM 40.7 GiB 52.5 GiB OOM — OOM
Llama 3.1 70B Instruct OOM 40.7 GiB 52.5 GiB OOM — OOM
Llama-3.3 Nemotron Super 49B OOM 28.7 GiB 36.7 GiB OOM — OOM
Mixtral 8x7B Instruct OOM 26.9 GiB 32.2 GiB OOM — OOM
Qwen3 32B OOM 18.9 GiB 28.3 GiB OOM — 11k
Qwen2.5 Coder 32B OOM 18.9 GiB 28.3 GiB OOM — 11k
DeepSeek-R1 Distill Qwen 32B OOM 18.9 GiB 28.3 GiB OOM — 11k
QwQ 32B Reasoning OOM 18.9 GiB 28.3 GiB OOM — 11k

Interactive NVIDIA RTX 4090 (24 GB) Simulator

27.4B Dense · GQA
1008 GB/s · 165.2 TFLOPS
3. Weight Quantization 4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 68k
Memory Allocation Footprint 20.17 / 23.00 GiB usable
Weights: 15.78 GiB KV Cache: 2.91 GiB Overhead: 1.49 GiB Headroom: 2.83 GiB
Decode Speed
41.2
tok/s (batch=1)
Prefill Speed
1888
prompt tok/s
TTFT Latency
2.21s
for 4,096 tok
Active Params
27.4B
of 27.4B total
llama.cpp Server Command 
llama-server -hf google/gemma-3-27b-it:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080
Context KV Cache Total VRAM Decode Status
4k 0.72 GiB 17.62 GiB 46.7 tok/s COMFORTABLE
8k 1.03 GiB 18.30 GiB 45.8 tok/s COMFORTABLE
16k 1.66 GiB 18.92 GiB 44.1 tok/s COMFORTABLE
32k 2.91 GiB 20.17 GiB 41.2 tok/s COMFORTABLE
64k 5.41 GiB 22.67 GiB 36.3 tok/s TIGHT
128k 10.41 GiB 27.67 GiB 29.4 tok/s OOM

Buy-vs-Cloud-vs-API Monthly TCO Break-Even

USE SERVERLESS API
Own Hardware/mo
$48.80
108.5h active + power
Cloud GPU/mo
$54.90
@$0.44/hr spot/pod
Serverless API/mo
$9.25
API cheaper

Recommended Hardware & Cloud Deployment

Contextual to fit status

Hardware FAQ — NVIDIA RTX 4090 (24 GB)

How much usable VRAM / GPU memory does NVIDIA RTX 4090 (24 GB) actually provide for LLMs?

NVIDIA RTX 4090 (24 GB) provides 23 GiB usable VRAM after subtracting 1.0 GiB driver/display overhead.

What is the largest Q4_K_M model NVIDIA RTX 4090 (24 GB) can run at 32k context?

With Q4_K_M weights and FP16 KV cache at 32k context, the largest fitting model in our verified catalog on NVIDIA RTX 4090 (24 GB) is Gemma 3 27B IT (27.4B total).