HARDWARE BENCHMARK & VRAM ENVELOPE • CONSUMER GPU24 GB VRAM • 936 GB/s
What Models Can Run on NVIDIA GeForce RTX 3090 (24GB GDDR6X)?
24 GB VRAM936 GB/s Bandwidth350W TDPMSRP: $1,499Used / eBay: ~$650Cloud: $0.29/hr on Vast.ai
NVIDIA GeForce RTX 3090 (24GB GDDR6X) provides 24 GB of usable memory with 936 GB/s memory bandwidth at 350W TDP. The #1 budget 24GB card for local LLMs, running 14B–32B models and Q4 70B quantized.
! PARTIAL OFFLOAD25.22 GB (105%)~7 t/s
Qwen 2.5 Coder 32B Instruct on NVIDIA GeForce RTX 3090 (24GB GDDR6X)
! Partial Offload — Slow CPU/RAM Spillover
! PARTIAL OFFLOAD
Required VRAM
25.22 GB
105% of hardware capacity (24 GB)Est. Decode Speed
~7 tok/s
Dense model @ 936 GB/sAvailable Headroom
-1.22 GB
Needs +1.22 GB more VRAMMemory Allocation & Headroom
25.22 GB allocated of 24 GB capacity
105% utilized
Weights: 19.5 GB
KV: (32K): 4.29 GB
Overhead: 1.43 GB
Deficit: -1.22 GB
Architecture & Throughput Transparency: Single-user decode speed (~7 tok/s) is memory-bandwidth bound at batch size = 1 on NVIDIA GeForce RTX 3090 (24GB GDDR6X) (936 GB/s). Prompt prefill / TTFT (~102 tok/s) is compute-bound and utilizes GPU Tensor cores. Actual latency varies with FlashAttention-2 and runtime engine (Ollama/vLLM/llama.cpp).
Out of VRAM? Don't downgrade model quality.
Out of VRAMSpin up an on-demand cloud GPU instance in 60 seconds with pre-installed Ollama and vLLM. (Qwen 2.5 Coder 32B Instruct @ 25.22 GB VRAM)
RunPod Cloud GPUsBest 1-Click Ollama & vLLM
$0.38/hr (RTX 4090 Spot) – $2.39/hr (H100)Vast.ai GPU MarketplaceCheapest Dual RTX 3090 / 4090
$0.18/hr (RTX 3060) – $0.78/hr (2x RTX 3090)Lambda GPU CloudEnterprise 8x H100 Clusters
$2.49/hr (H100) – $19.92/hr (8x H100)Quantization Precision Matrix
32K ctx • KV: Q8_0| Quantization | Weights | KV Cache | Total VRAM | Est. Speed | Fit Status |
|---|---|---|---|---|---|
| FP1616 bpw | 65 GB | 4.29 GB | 72.32 GB (301%) | 0 t/s | OOM |
| Q8_08.5 bpw | 34.45 GB | 4.29 GB | 40.7 GB (170%) | 0 t/s | OOM |
| Q6_K6.56 bpw | 26.65 GB | 4.29 GB | 32.62 GB (136%) | 0 t/s | OOM |
| Q5_K_M5.68 bpw | 23.07 GB | 4.29 GB | 28.92 GB (121%) | ~6 t/s | PARTIAL |
| Q4_K_M4.8 bpw | 19.5 GB | 4.29 GB | 25.22 GB (105%) | ~7 t/s | PARTIAL |
| FP44.25 bpw | 17.23 GB | 4.29 GB | 22.87 GB (95%) | ~33 t/s | FITS |
| IQ3_XS3.3 bpw | 13.32 GB | 4.29 GB | 18.83 GB (78%) | ~42 t/s | FITS |
All Open-Weights Models Tested on NVIDIA GeForce RTX 3090 (24GB GDDR6X) (Q4_K_M @ 32K Context)
Standard Q4_K_M precision with 32K context window and Q8_0 KV cache
| Model | Architecture | Total VRAM Needed | Headroom | Est. Decode Speed | Compatibility | Detailed Config |
|---|---|---|---|---|---|---|
| Qwen 2.5 Coder 32B Instruct Alibaba Cloud | 32.5B Dense | 25.22 GB(105%) | -1.22 GB | ~7 tok/s | ! PARTIAL | Configure CLI → |
| DeepSeek-R1 (671B MoE) DeepSeek | 671B MoE | 427.26 GB(999%) | -403.26 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| DeepSeek-V3 (671B MoE) DeepSeek | 671B MoE | 427.26 GB(999%) | -403.26 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Meta Llama 3.3 70B Instruct Meta AI | 70.6B Dense | 49.96 GB(208%) | -25.96 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| DeepSeek-R1 Distill Qwen 32B DeepSeek | 32.8B Dense | 25.41 GB(106%) | -1.41 GB | ~6 tok/s | ! PARTIAL | Configure CLI → |
| DeepSeek-R1 Distill Qwen 14B DeepSeek | 14.8B Dense | 13.16 GB(55%) | +10.84 GB | ~63 tok/s | ✓ FITS | Configure CLI → |
| Qwen 2.5 72B Instruct Alibaba Cloud | 72.7B Dense | 51.27 GB(214%) | -27.27 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Meta Llama 3.1 8B Instruct Meta AI | 8B Dense | 7.87 GB(33%) | +16.13 GB | ~114 tok/s | ✓ FITS | Configure CLI → |
| Microsoft Phi-4 14B Microsoft | 14.7B Dense | 13.24 GB(55%) | +10.76 GB | ~63 tok/s | ✓ FITS | Configure CLI → |
| DeepSeek Coder V2 236B MoE DeepSeek | 236B MoE | 147.91 GB(616%) | -123.91 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Mistral Large 2 (123B Instruct) Mistral AI | 123B Dense | 83.04 GB(346%) | -59.04 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Meta Llama 3.1 405B Instruct Meta AI | 405B Dense | 260.72 GB(999%) | -236.72 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Google Gemma 2 27B Instruct Google | 27.2B Dense | 19.18 GB(80%) | +4.82 GB | ~36 tok/s | ✓ FITS | Configure CLI → |
| Google Gemma 2 9B Instruct Google | 9.2B Dense | 7.87 GB(33%) | +16.13 GB | ~104 tok/s | ✓ FITS | Configure CLI → |
| Qwen 3.8 27B Instruct Alibaba Cloud | 27.2B Dense | 21.93 GB(91%) | +2.07 GB | ~35 tok/s | • TIGHT FIT | Configure CLI → |
| Qwen 3.8 235B-A22B MoE Alibaba Cloud | 235B MoE | 149.84 GB(624%) | -125.84 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Qwen 3.6 35B-A3B MoE Alibaba Cloud | 35B MoE | 24.1 GB(100%) | -0.1 GB | ~61 tok/s | ! PARTIAL | Configure CLI → |
| DeepSeek-V4.1-Flash (552B Engram MoE) DeepSeek | 552B MoE | 343.85 GB(999%) | -319.85 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Meta Muse Spark Open (816B-A42B MoE) Meta AI | 816B MoE | 512.86 GB(999%) | -488.86 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Google Gemma 4 26B MoE Google | 26.4B MoE | 20.23 GB(84%) | +3.77 GB | ~185 tok/s | ✓ FITS | Configure CLI → |
| Google Gemma 4 E4B (Edge 4B) Google | 4.1B Dense | 4.24 GB(18%) | +19.76 GB | ~226 tok/s | ✓ FITS | Configure CLI → |
| Meta Llama 4 Scout (109B-A17B MoE) Meta AI | 109B MoE | 71.66 GB(299%) | -47.66 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| Microsoft Phi-4 Mini 3.8B Microsoft | 3.8B Dense | 3.92 GB(16%) | +20.08 GB | ~245 tok/s | ✓ FITS | Configure CLI → |
| Mistral Small 3 24B Instruct Mistral AI | 24B Dense | 18.33 GB(76%) | +5.67 GB | ~40 tok/s | ✓ FITS | Configure CLI → |
| DeepSeek-R1 Distill Llama 70B DeepSeek | 70.6B Dense | 49.96 GB(208%) | -25.96 GB | 0 (OOM) | ✕ OOM | Configure CLI → |
| DeepSeek-R1 Distill Qwen 7B DeepSeek | 7.6B Dense | 6.41 GB(27%) | +17.59 GB | ~127 tok/s | ✓ FITS | Configure CLI → |
| Qwen 2.5 Coder 7B Instruct Alibaba Cloud | 7.6B Dense | 6.41 GB(27%) | +17.59 GB | ~127 tok/s | ✓ FITS | Configure CLI → |