How we calculate memory requirements

Parameters (billions) × 1 GB per billion params × 1.2 overhead factor = Minimum GPU memory needed

Why 1 GB per billion? When stored at 16-bit (fp16/bf16) precision — the standard for inference — each parameter takes 2 bytes. One billion parameters = 2 GB. We use 1 GB as a conservative per-billion estimate for 8-bit quantized loading, which most open-source tools support. At full fp16, multiply by 2.

Why 1.2× overhead? The model weights aren't the only thing in GPU memory. The KV cache (storing attention state during generation), activations, and runtime buffers add roughly 15–25% on top of the weights themselves. We use 1.2× as a minimum — production workloads often budget 1.5–2×.

Z.ai (zai-org) MIT License Arena.ai Leaderboard ✓

GLM-5.2 (Max)

One of the largest open-source language models available, ranked #10 on the Arena Agent leaderboard. Mixture-of-Experts architecture with 40B active params per token. Excels at complex reasoning, code, and multilingual tasks.

Memory Calculation

Step 1: Parameter count
753 billion parameters
Verified from the official model card on HuggingFace and the Arena.ai leaderboard (open-source, MIT/Apache 2.0 filter).
×
Step 2: Memory per billion params
1 GB / billion params
8-bit quantization baseline. At full fp16 precision, use 2 GB / billion instead.
=
Weights only
753 GB
×
Step 3: Overhead factor
1.2×
KV cache + activations + runtime buffers add ~20% minimum.
=
Minimum GPU memory needed
904 GB
This is the floor. Production deployments often add 50–100% more for concurrency.
753 B × 1 GB × 1.2 = 904 GB minimum

Minimum Hardware to Run This Model

MINIMUM VIABLE

NVIDIA H200 HGX 8-GPU Server

1,128 GB available ✓ 224 GB headroom

A single 8×H200 server (1,128 GB) covers the 904 GB needed with ~224 GB headroom.

Price $370,000
Power 5,600 W
Form factor 8-GPU rack server (fits 4U rack space)
View full specs →
Why can't I run this on a cheaper card?
  • RTX 5090 (32 GB): Would need 29 cards, each with isolated memory. Can't pool across consumer cards reliably. ✗
  • RTX PRO 6000 (96 GB): Would need 10 cards. Multi-card workstation inference works with tensor parallelism but adds latency. ✗ for production
  • 8× H200 HGX (1,128 GB): 904 GB needed, 1,128 GB available. All memory shared via NVLink. ✅

Tencent Apache 2.0 License Arena.ai Leaderboard ✓

Tencent Hy3

Tencent's open frontier model. Strong performance on Chinese and English benchmarks, designed for enterprise deployment.

Memory Calculation

Step 1: Parameter count
295 billion parameters
Verified from the official model card on HuggingFace and the Arena.ai leaderboard (open-source, MIT/Apache 2.0 filter).
×
Step 2: Memory per billion params
1 GB / billion params
8-bit quantization baseline. At full fp16 precision, use 2 GB / billion instead.
=
Weights only
295 GB
×
Step 3: Overhead factor
1.2×
KV cache + activations + runtime buffers add ~20% minimum.
=
Minimum GPU memory needed
354 GB
This is the floor. Production deployments often add 50–100% more for concurrency.
295 B × 1 GB × 1.2 = 354 GB minimum

Minimum Hardware to Run This Model

MINIMUM VIABLE

NVIDIA H200 HGX 8-GPU Server

1,128 GB available ✓ 774 GB headroom

A single 8×H200 server (1,128 GB) covers the 354 GB needed with 3× headroom — great throughput.

Price $370,000
Power 5,600 W
Form factor 8-GPU rack server (fits 4U rack space)
View full specs →
Why can't I run this on a cheaper card?
  • RTX 5090 (32 GB): Would need 12 cards, each with isolated memory. Can't pool across consumer cards reliably. ✗
  • RTX PRO 6000 (96 GB): Would need 4 cards. Multi-card workstation inference works with tensor parallelism but adds latency. ✗ for production
  • 8× H200 HGX (1,128 GB): 354 GB needed, 1,128 GB available. All memory shared via NVLink. ✅

DeepSeek MIT License Arena.ai Leaderboard ✓

DeepSeek V4-Flash

DeepSeek's efficient frontier model. Delivers frontier-level reasoning at lower latency than full-scale MoE models, making it practical for production inference.

Memory Calculation

Step 1: Parameter count
284 billion parameters
Verified from the official model card on HuggingFace and the Arena.ai leaderboard (open-source, MIT/Apache 2.0 filter).
×
Step 2: Memory per billion params
1 GB / billion params
8-bit quantization baseline. At full fp16 precision, use 2 GB / billion instead.
=
Weights only
284 GB
×
Step 3: Overhead factor
1.2×
KV cache + activations + runtime buffers add ~20% minimum.
=
Minimum GPU memory needed
341 GB
This is the floor. Production deployments often add 50–100% more for concurrency.
284 B × 1 GB × 1.2 = 341 GB minimum

Minimum Hardware to Run This Model

MINIMUM VIABLE

NVIDIA H200 HGX 8-GPU Server

1,128 GB available ✓ 787 GB headroom

A single 8×H200 server (1,128 GB) covers the 341 GB needed. Strong choice for a startup that wants cutting-edge open-source inference.

Price $370,000
Power 5,600 W
Form factor 8-GPU rack server (fits 4U rack space)
View full specs →
Why can't I run this on a cheaper card?
  • RTX 5090 (32 GB): Would need 11 cards, each with isolated memory. Can't pool across consumer cards reliably. ✗
  • RTX PRO 6000 (96 GB): Would need 4 cards. Multi-card workstation inference works with tensor parallelism but adds latency. ✗ for production
  • 8× H200 HGX (1,128 GB): 341 GB needed, 1,128 GB available. All memory shared via NVLink. ✅