Deploy GPU ↗
HARDWARE FIT CHECK • DeepSeek Coder V2 236B MoE + NVIDIA GeForce RTX 3090 (24GB GDDR6X)147.91 GB / 24 GB VRAM

Can I Run DeepSeek Coder V2 236B MoE on NVIDIA GeForce RTX 3090 (24GB GDDR6X)?

Full VRAM footprint analysis, tokens/sec generation benchmark, and copy-paste Ollama & vLLM runtime commands for running DeepSeek Coder V2 236B MoE (236B) on NVIDIA GeForce RTX 3090 (24GB GDDR6X) (24 GB).

✕ OOM147.91 GB (616%)0 t/s

Target Hardware & Model

interactive
338-language coding & repo-level context for 64GB–128GB unified memory
DeepSeek Coder V2 236B MoE on NVIDIA GeForce RTX 3090 (24GB GDDR6X)MoE (16B Active / 236B Total)

✕ Out of Memory — Exceeds VRAM Capacity

✕ OOM
Required VRAM
147.91 GB
616% of hardware capacity (24 GB)
Est. Decode Speed
0 tok/s
16B active params @ 936 GB/s
Available Headroom
-123.91 GB
Needs +123.91 GB more VRAM
Memory Allocation & Headroom
147.91 GB allocated of 24 GB capacity
616% utilized
Weights: 141.6 GBMoE (16B Active)
KV: (32K): 0.6 GBMLA (-85%)
Overhead: 5.71 GB
Deficit: -123.91 GB
Architecture & Throughput Transparency: Single-user decode speed (~0 tok/s) is memory-bandwidth bound at batch size = 1 on NVIDIA GeForce RTX 3090 (24GB GDDR6X) (936 GB/s). Prompt prefill / TTFT (~0 tok/s) is compute-bound and utilizes GPU Tensor cores. Actual latency varies with FlashAttention-2 and runtime engine (Ollama/vLLM/llama.cpp).
MoE Memory vs Speed Architecture: While decode throughput scales with the active parameter count (16B parameters streamed per token), all 236B parameters must be resident in VRAM so the dynamic router can route tokens to any expert layer.
Multi-Head Latent Attention (MLA): DeepSeek compresses Key-Value projections into a low-rank latent vector, cutting KV memory footprint by ~85% compared to standard MHA/GQA.

Local Hardware Configurations & Pricing

Out of VRAM: High-Capacity Hardware Upgrades
High-bandwidth VRAM specs and current market prices
Multi-GPU Rig↑ Upgrade
48 GB VRAM

2x NVIDIA RTX 3090 Dual Rig (48GB Total VRAM)

Best price-to-VRAM homelab rig for Llama 3.3 70B Q4_K_M & Qwen 2.5 Coder 32B Q8_0

1648 GB/s700W TDP
$1,550Amazon Prime →
Apple Silicon↑ Upgrade
54 GB VRAM

Apple Mac Mini M4 Pro (64GB Unified Memory)

Silent 65W desktop running 32B–70B models with large KV context

273 GB/s65W TDP
$1,999Amazon Prime →
Apple Silicon↑ Upgrade
112 GB VRAM

Apple Mac Studio M4 Max (128GB Unified Memory)

Running Mistral Large 2 123B & Llama 3.3 70B Q8_0 locally with 128K context

546 GB/s145W TDP
$3,499Amazon Prime →
Consumer GPU
24 GB VRAM

NVIDIA GeForce RTX 3090 (24GB GDDR6X)

The #1 budget 24GB card for local LLMs, running 14B–32B models and Q4 70B quantized

936 GB/s350W TDP
$1,499Amazon Prime →

Out of VRAM? Don't downgrade model quality.

Out of VRAM

Spin up an on-demand cloud GPU instance in 60 seconds with pre-installed Ollama and vLLM. (DeepSeek Coder V2 236B MoE @ 147.91 GB VRAM)

RunPod Cloud GPUsBest 1-Click Ollama & vLLM
$0.38/hr (RTX 4090 Spot) – $2.39/hr (H100)
Launch 1-Click GPU Pod →
Vast.ai GPU MarketplaceCheapest Dual RTX 3090 / 4090
$0.18/hr (RTX 3060) – $0.78/hr (2x RTX 3090)
Rent Spot GPU →
Lambda GPU CloudEnterprise 8x H100 Clusters
$2.49/hr (H100) – $19.92/hr (8x H100)
Check H100 Availability →

Quantization Precision Matrix

32K ctx • KV: Q8_0
QuantizationWeightsKV CacheTotal VRAMEst. SpeedFit Status
FP1616 bpw472 GB0.6 GB489.87 GB (999%)0 t/sOOM
Q8_08.5 bpw250.16 GB0.6 GB260.27 GB (999%)0 t/sOOM
Q6_K6.56 bpw193.52 GB0.6 GB201.64 GB (840%)0 t/sOOM
Q5_K_M5.68 bpw167.56 GB0.6 GB174.77 GB (728%)0 t/sOOM
Q4_K_M4.8 bpw141.6 GB0.6 GB147.91 GB (616%)0 t/sOOM
FP44.25 bpw125.08 GB0.6 GB130.81 GB (545%)0 t/sOOM
IQ3_XS3.3 bpw96.76 GB0.6 GB101.5 GB (423%)0 t/sOOM
FAQ & Reference

Frequently Asked Questions: DeepSeek Coder V2 236B MoE on NVIDIA GeForce RTX 3090 (24GB GDDR6X)

How much VRAM does DeepSeek Coder V2 236B MoE need on NVIDIA GeForce RTX 3090 (24GB GDDR6X)?

At standard 4-bit quantization (Q4_K_M) with 32K context, DeepSeek Coder V2 236B MoE requires 147.91 GB of total VRAM (weights + KV cache + CUDA runtime overhead).

What inference speed will DeepSeek Coder V2 236B MoE achieve on NVIDIA GeForce RTX 3090 (24GB GDDR6X)?

Based on NVIDIA GeForce RTX 3090 (24GB GDDR6X)'s 936 GB/s memory bandwidth, expected decode speed is approximately 0 tokens per second.

Can I expand context beyond 32K tokens?

Each additional 32K tokens of context requires additional KV cache memory. For longer context windows, consider 8-bit or 4-bit KV cache quantization.