Kimi K3 (2.8T-A104B) on RTX 4090 D (24GB) — VRAM, KV Cache & Throughput
Kimi K3 (2.8T-A104B) (2780B total, 104B active, Multi-head Latent Attention (MLA) MoE) uses 1570.1 GiB of 4-bit weights (Q4_K_M), requiring 1571.49 GiB total at 8K context and 1573.94 GiB at 32K on RTX 4090 D (~0.9 tok/s decode).
Exceeds Available GPU Memory
Config-derived formula
HF config.json ↗
Required GPU Memory
1571.49 GiB
1687.37 GB (decimal)
Usable Hardware Memory
22.8 GiB
24 GB (physical)
Safe Headroom
-1548.69 GiB
6548% utilization
Est. Decode Throughput
~0.9 tok/s
104B active params
Est. Prefill Throughput
~363 tok/s
4K Prompt TTFT: ~11299 ms
Est. Cost / 1M Tokens
$9382.72
Multi-head Latent Attention (MLA) MoE · Custom
Recommended Hardware for Current Configuration
Best Single GPU
Multi-GPU Required
—
Best Value / Used GPU
—
—
Apple Unified Memory
Exceeds 512 GB
—
Cloud From
DigitalOcean · $30.40 / hr
8x Instinct MI325X 256GB
Production Launch Commands
llama.cpp (GGUF)
llama-server -hf unsloth/Kimi-K3-GGUF:Q4_K_M \ --ctx-size 8192 -ngl 999 --flash-attn
vLLM
vllm serve moonshotai/Kimi-K3 \ --max-model-len 8192 --gpu-memory-utilization 0.92
Live Token Speed Simulator (TTFT ~11299ms · ~0.9 tok/s)
[Kimi K3 (2.8T-A104B) · 1571.49 GiB] allocated on target GPU. Prefill completed in TTFT window; streaming autoregressive tokens at memory-bandwidth-bound rate...
Cloud GPU Comparison & Rent-vs-Buy TCO
Verified public rates across RunPod, Vast.ai, Vultr, and DigitalOcean.
Purchase break-even
~0.3 months
Est. local power: $9.04 / mo
| Provider | Recommended Instance | Total VRAM | Hourly Rate | Monthly at Selected Hours | Billing Mode | Deploy |
|---|---|---|---|---|---|---|
| RunPod | 8x B300 288GB Blackwell Ultra (Community Cloud) | 2304 GB | $55.52 / hr | $6751.23 / mo | Per-second | Deploy GPU ↗ |
| Vast.ai | — | — | — | — | Spot / On-demand | Deploy GPU ↗ |
| Vultr ($300 Free Credit*) | — | — | — | — | Hourly | Deploy GPU ↗ |
| Vultr Cloud GPU | — | — | — | — | Hourly | Deploy GPU ↗ |
| DigitalOcean | 8x Instinct MI325X 256GB (Dedicated Instance) | 2048 GB | $30.40 / hr | $3696.64 / mo | Per-second | Deploy GPU ↗ |
Rates reflect public listings on the verification date and fluctuate with regional inventory.
Fine-Tuning VRAM Requirements (QLoRA / LoRA / Full) — Kimi K3 (2.8T-A104B)
QLoRA (4b)
Exceeds Available GPU Memory
1585.83 GiB
Model Weights: 1570.1 GiB · Gradients + Optimizer: 14.56 GiB · Act: 1.17 GiB
Recommended GPU: Multi-GPU Required
LoRA (16b)
Exceeds Available GPU Memory
5202.6 GiB
Model Weights: 5178.15 GiB · Gradients + Optimizer: 22.88 GiB · Act: 1.57 GiB
Recommended GPU: Multi-GPU Required
Full BF16
Exceeds Available GPU Memory
21233.18 GiB
Model Weights: 5178.15 GiB · Gradients + Optimizer: 16052.26 GiB · Act: 2.77 GiB
Recommended GPU: Multi-GPU Required
Engineering Sizing Notes & Optimization Playbook
Quick Hardware Verdict
A single RTX 4090 D (24GB) cannot hold 4-bit Kimi K3 (2.8T-A104B) (1571.49 GiB at 8K) entirely in VRAM. Use multi-GPU tensor parallelism, offload layers to system RAM, or deploy on a larger cloud GPU.
KV Cache & Runtime Tuning
Kimi K3 (2.8T-A104B) uses Multi-head Latent Attention (MLA) MoE (2780B total, 104B active per token). Moving from 8K to 32K context expands FP16 KV cache from 0.82 GiB to 3.27 GiB; enabling 8-bit KV cache (llama.cpp -ctk q8_0 -ctv q8_0 or vLLM --kv-cache-dtype fp8) saves 1.64 GiB at 32K (reducing total VRAM to 1572.3 GiB).
VRAM Scaling Across Context Windows (4K to 128K, 4-bit Weights)
Quantization Precision & Memory Matrix — Kimi K3 (2.8T-A104B) @ 8K
| Quantization Format | Weights (GiB) | KV Cache | Total VRAM (GiB) | Decode Speed | Compatibility (RTX 4090 D) | Recommended GPU | Data Source |
|---|---|---|---|---|---|---|---|
| Native Weights (BF16 / FP8 / MXFP4 — Lossless) (MXFP4) | 5178.15 GiB | 0.82 GiB | 5179.54 GiB (5561.49 GB) | ~0.8 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Official checkpoint size |
| FP8 / GGUF Q8_0 (8-bit — Near-Lossless) (Q8_0) | 2750.89 GiB | 0.82 GiB | 2752.28 GiB (2955.24 GB) | ~0.8 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| GGUF Q6_K (6-bit — High Precision) (Q6_K) | 2123.04 GiB | 0.82 GiB | 2124.43 GiB (2281.09 GB) | ~0.8 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| GGUF Q5_K_M (5-bit — Balanced) (Q5_K_M) | 1841.72 GiB | 0.82 GiB | 1843.11 GiB (1979.02 GB) | ~0.8 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| Q4_K_M / UD-Q4_K_XL / AWQ (4-bit — Recommended) (Q4_K_M) | 1570.1 GiB | 0.82 GiB | 1571.49 GiB (1687.37 GB) | ~0.9 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| GGUF IQ4_XS (4.25-bit — Compact) (IQ4_XS) | 1375.79 GiB | 0.82 GiB | 1377.18 GiB (1478.74 GB) | ~1 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| GGUF Q3_K_M (3-bit — Low VRAM) (Q3_K_M) | 1265.71 GiB | 0.82 GiB | 1267.1 GiB (1360.54 GB) | ~1.1 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
| GGUF UD-Q2_K_XL (2-bit — Extreme MoE) (UD-Q2_K_XL) | 890.47 GiB | 0.82 GiB | 891.86 GiB (957.63 GB) | ~1.6 tok/s | Exceeds Available GPU Memory | Multi-GPU Required | Config-derived formula |
Technical Specifications & FAQ
Does Kimi K3 (2.8T-A104B) fit on RTX 4090 D (24GB)?
In 4-bit (Q4_K_M) at 8,192 context, Kimi K3 (2.8T-A104B) requires 1571.49 GiB total GPU memory (1570.1 GiB weights + 0.82 GiB KV cache + 0.57 GiB runtime buffers) against 22.8 GiB usable on RTX 4090 D, delivering ~0.9 tok/s decode.
How much VRAM does Kimi K3 (2.8T-A104B) need at 32K context?
At 32,768 context tokens, Kimi K3 (2.8T-A104B) (Multi-head Latent Attention (MLA) MoE) allocates 3.27 GiB for KV cache and state, bringing total 4-bit VRAM to 1573.94 GiB (1690.01 GB decimal).
Evaluate This Model on Other GPUs
RTX 3060 12GB (12GB)RTX 4060 Ti 16GB (16GB)RTX 5060 Ti 16GB (16GB)RTX 4070 Ti Super (16GB)RTX 4080 Super (16GB)RTX 5070 12GB (12GB)RTX 5070 Ti (16GB)RTX 5080 (16GB)RTX 3090 (24GB)RTX 4090 (24GB)RTX 5090 (32GB)RTX 5090 D (32GB)RTX A6000 (48GB)RTX 6000 Ada (48GB)RTX PRO 6000 Blackwell (96GB)DGX Spark (GB10 128GB) (128GB)L4 24GB (24GB)A40 48GB (48GB)L20 48GB (48GB)L40S 48GB (48GB)A800 80GB SXM (80GB)A100 80GB SXM (80GB)H800 80GB SXM (80GB)H100 80GB SXM (80GB)H20 96GB HBM3 (96GB)H20 141GB HBM3e (141GB)H200 141GB SXM (141GB)B200 180GB (180GB)B300 288GB Blackwell Ultra (288GB)Ascend 910B 64GB (昇腾) (64GB)Ascend 910C 128GB (昇腾) (128GB)Atlas 300I Duo 96GB (310P) (96GB)MLU590 80GB (寒武纪思元) (80GB)DCU K100-AI 64GB (海光深算) (64GB)MetaX C500 64GB (沐曦曦云) (64GB)Kunlunxin P800 64GB (昆仑芯) (64GB)MTT S4000 48GB (摩尔线程) (48GB)Radeon RX 7900 XTX 24GB (24GB)Radeon AI PRO R9700 32GB (32GB)Ryzen AI Max+ 395 128GB (128GB)Instinct MI300X 192GB (192GB)Instinct MI325X 256GB (256GB)Instinct MI355X 288GB (288GB)Arc Pro B60 24GB (24GB)Gaudi 3 128GB HBM2e (128GB)Mac mini M4 Pro 64GB (64GB)MacBook Pro M4 Max 128GB (128GB)Mac Studio M4 Ultra 256GB (256GB)Mac Studio M3 Ultra 512GB (512GB)
Evaluate Other Models on This Hardware
Qwen3.8 Flash Next (125B-A6B)Qwen3.8 27BQwen3.8 2.4T-A95BQwen3.6 35B-A3BQwen3.5 397B-A17BQwen3.5 122B-A10BQwen3.5 35B-A3BQwen3.5 27BQwen3.5 9BQwen3.5 4BQwen3 Coder 480B-A35BQwen3 Coder Next 80B-A3BQwen3 Coder 30B-A3BQwen3 30B-A3B Instruct 2507Qwen3 235B-A22B Instruct 2507Qwen3 32BQwen3 14BQwen3 8BQwen3 4B Instruct 2507Qwen3 1.7BQwen3 0.6BQwen2.5 72B InstructQwen2.5 Coder 32B InstructQwQ 32B ReasoningDeepSeek V4.1 Flash (552B-A16B)DeepSeek V4 FlashDeepSeek V3.2 (685B)DeepSeek R1-0528 671BDeepSeek V3-0324 671BDeepSeek R1 Distill Llama 70BDeepSeek R1 Distill Qwen 32BDeepSeek R1 Distill Qwen 14BDeepSeek R1 Distill Llama 8Bgpt-oss-20bgpt-oss-120bGemma 4 31B ITGemma 4 26B-A4B ITGemma 4 12B ITGemma 4 E4B ITGemma 4 E2B ITGemma 3 27B ITGemma 3 12B ITGemma 3 4B ITGLM-5.3 Flash (320B-A18B)GLM-5.3 (753B)GLM-5 (744B)GLM-4.5 (355B-A32B)GLM-4.5 Air (110B-A12B)GLM-4 32B-0414Kimi K3.1 Pro (2.8T-A104B)Kimi K2 Instruct (1T-A32B)Kimi-Dev 72BMiniMax M3.1 Flash (428B-A23B)MiniMax M2.5 (229B-A10B)Step-5 Preview (600B-A27B)Step-3.7 Flash (198B-A11B)Step-3.5 Flash (197B-A11B)ERNIE 4.5 300B-A47B (文心)ERNIE 4.5 21B-A3B (文心)Hunyuan A13B Instruct (混元 80B)MiniCPM5 2B (面壁)Llama 4 Maverick 17B-128E (400B)Llama 4 Scout 17B-16E (109B)Llama 3.1 405B InstructLlama 3.3 70B InstructLlama 3.1 8B InstructLlama 3.2 3B InstructMistral Large 4 (1T-A52B)Mistral Large 2 123BMixtral 8x22B InstructMixtral 8x7B InstructMistral Small 3.2 24BDevstral Small 24BNemotron 3 Ultra 550B-A55BNemotron 3 Super 120B-A12BNemotron 3 Nano 30B-A3BCohere Command A 111BPhi-4 14BLiquid LFM3 1.7B