Sourced from the Arena Agent Leaderboard

Three real open-source models

Pulled from arena.ai/leaderboard/agent, filtered to models whose license is genuinely open — MIT or Apache 2.0, not a "free to look at" custom license. Every parameter count and memory requirement below is calculated, not asserted.

Apache 2.0

Gemma 4 31B

A dense (not mixture-of-experts) 31-billion-parameter model. Proof that a smaller, far cheaper-to-run open model can still be frontier-competitive — this is usually the right starting point if you don't need the absolute largest model available.

Google · Rank #31 on the Arena Agent Leaderboard (arena.ai/leaderboard/agent) — the first Gemma generation Google has released under Apache 2.0.

The memory math, shown in full: 31B parameters (dense) × 2 GB/B (FP16/BF16 inference) = 61 GB baseline, + 20% for KV cache / context overhead = 74 GB minimum GPU memory required.
See the startup build → See the mid-size build →
MIT

GLM-5.2

A 744-billion-parameter mixture-of-experts model — only about 40 billion of those parameters are active for any single response, which keeps it fast, but all 744 billion still have to be loaded into GPU memory at once since any token can route to any expert.

Z.ai (Zhipu AI) · Rank #8 on the Arena Agent Leaderboard — the highest-ranked fully open-weight model on the board.

The memory math, shown in full: 744B total parameters (mixture-of-experts, ~40B active per token — but all 744B must be resident in memory) × 2 GB/B (FP16/BF16 inference) = 1,488 GB baseline, + 20% for KV cache / context overhead = 1,786 GB minimum GPU memory required.
See the startup build → See the mid-size build →
MIT

DeepSeek V4 Pro

A 1.6-trillion-parameter mixture-of-experts model — the largest model in our lineup. It's a genuine frontier-scale system, and the hardware required to run it reflects that honestly: even the 'minimum' setup is a serious multi-server investment.

DeepSeek · Rank #21 on the Arena Agent Leaderboard.

The memory math, shown in full: 1,600B total parameters (mixture-of-experts, ~49B active per token — but all 1,600B must be resident in memory) × 2 GB/B (FP16/BF16 inference) = 3,200 GB baseline, + 20% for KV cache / context overhead = 3,840 GB minimum GPU memory required.
See the startup build → See the mid-size build →