AI models we advise on

We picked exactly three models from the world agent leaderboard (arena.ai/leaderboard/agent): all genuinely open-source (MIT or Modified MIT license) and all very large (well over 100 billion parameters). For each, we show the minimum hardware needed to run it, with the arithmetic done in the open.

Kimi K2.6

Moonshot AI · leaderboard rank #18 · Modified MIT license

1100B total parameters

32B active per token (Mixture-of-Experts)

The official Hugging Face model card (huggingface.co/moonshotai/Kimi-K2.6) shows two figures: the architecture table states "Total Parameters: 1T" (1,000 billion), while the file-metadata sidebar — measured directly from the actual model weight files — states "Model size: 1.1T params" (1,100 billion).

We use the 1,100 billion figure, the same way this store prefers GLM-5.2's directly-stated "Model size" figure over a rounder secondary number.

The 32B active-parameter figure comes from the architecture table, which is the only place it's stated, with the license listed as "modified-mit".

GLM-5.2

Z.ai (Zhipu AI) · leaderboard rank #7 · MIT license

753B total parameters

39B active per token (Mixture-of-Experts)

The official Hugging Face model card (huggingface.co/zai-org/GLM-5.2) states "Model size: 753B params." Some secondary sources cite 744B instead — we use the 753B figure stated directly on the model card.

The ~39B active-parameter figure comes from third-party technical write-ups rather than being stated as plainly on the card itself, so treat it as less certain than the 753B total.

DeepSeek V4 Flash

DeepSeek · leaderboard rank #17 · MIT license

284B total parameters

13B active per token (Mixture-of-Experts)

The official Hugging Face model card (huggingface.co/deepseek-ai/DeepSeek-V4-Flash) states "284B parameters (13B activated)" and lists the license as MIT.