Model Weights & Local Inference Memory Estimator

LLM Sizing v1.0

VRAM & Memory Allocation Profile

Target: 2x RTX 3090 (48 GB Total Capacity)
PASS - FIT IN VRAM
38.5 GB
4.0 GB
0.0 GB
5.5 GB Free
Model Weights: 38.5 GB
KV Cache Overhead: 4.0 GB
CUDA Context/Compute: 0.0 GB
VRAM Overflow: 0.0 GB
Total Memory Required
42.5 GB
88.5% of hardware limit
Est. Generation Speed
44.0 tok/s
Bandwidth limited
Time To First Token
112 ms
Prompt processing
Single 24GB GPU Run
No
Requires multi-GPU/Offload

VRAM Breakdown by Category

KV Cache Scaling vs Context Window

Hardware Compatibility Analysis

Recommended Hardware Sizing 2x RTX 3090 (48GB) or Mac Studio M2/M3 Ultra (64GB+)
Single 24GB VRAM GPU Compatibility false
Effective Bits Per Weight 4.50 bits
KV Cache Size (FP16 Attention Precision) 4.0 GB
Proof Assessment Record:
Model: Kimi K3 / Frontier Open Weights
Total VRAM Required: 42.5 GB
Weights Memory: 38.5 GB
KV Cache VRAM: 4.0 GB
Single 24GB GPU Compatible: false
Recommended Hardware: 2x RTX 3090 (48GB) or Mac Studio M2/M3 Ultra (64GB+)

Read the written weight and cache estimator assumptions

Read the explanation

The executable sizing formula multiplies parameter billions by assigned bits per weight and divides by eight, then rounds to one decimal gigabyte. For seventy billion parameters, FP16 gives one hundred forty gigabytes and the source Q4 K M assignment of four point five bits gives thirty-nine point four. Both bars use three pixels per decimal gigabyte. A source comment mentions an overhead factor, but the executable weight expression does not multiply by it. These are written assumptions, not measured model allocations. This source normalizes cache per token by parameter count, not explicit layer and head inputs. At seventy billion, thirty-two thousand seven hundred sixty-eight tokens and batch one, its written formula gives a quarter gigabyte before rounding, displayed as zero point three. Doubling batch gives zero point five. Both bars show those displayed rounded terms at eight hundred pixels per gigabyte. The initial markup and comments mention four gigabytes, but that is not what this executable formula computes. This is a source-specific approximation, not actual KV geometry. Runtime overhead is explicitly weight memory times zero, so it adds nothing. With default Q4 weight thirty-nine point four and rounded cache zero point three, total is thirty-nine point seven gigabytes. A forty-eight gigabyte assigned hardware capacity leaves eight point three. Both bars use ten pixels per gigabyte. Throughput is an assigned bandwidth divided by weights and latency another formula. They are estimates, not current hardware benchmarks or product recommendations.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.