N Natalie's AI Hardware Co.
Model library

Open-Source AI Models

We only list models that (1) are genuinely open-source under an Both are "permissive" open-source licenses: anyone can download, modify, and even sell products built on the model, with minimal restrictions. That's different from many AI models which use custom licenses that restrict commercial use, require revenue sharing, or cap the number of users. We only stock models under these two license types. — not a custom or restricted license, (2) have roughly 100 billion parameters or more, and (3) appear on the LMArena leaderboard or are widely recognized flagship releases. We verified each license directly against its official Hugging Face model card or license file.

The memory formula
The number of learned numeric values ("weights") inside an AI model. Roughly speaking, more parameters means a more capable but more memory-hungry model. "120B" means 120 billion parameters. determine memory: (parameters in billions × 1 GB) + 20% overhead. The 20% covers the runtime, active requests, and context memory. See In real deployments, memory needs also depend on numeric precision (e.g. FP16 vs INT8 vs INT4), which changes bytes-per-parameter. This store uses a simplified, transparent formula so beginners can follow the logic: (parameters in billions × 1 GB) + 20% overhead for the runtime, active requests, and context memory. for why this is intentionally simplified.

gpt-oss-120b

OpenAI · Mixture-of-Experts (MoE) · Apache-2.0

117B
Total parameters117B (5.1B active per token)
Memory required140.4 GB

Mixtral 8x22B

Mistral AI · Mixture-of-Experts (MoE) · Apache-2.0

141B
Total parameters141B (39B active per token)
Memory required169.2 GB

Qwen3-235B-A22B

Alibaba · Mixture-of-Experts (MoE) · Apache-2.0

235B
Total parameters235B (22B active per token)
Memory required282.0 GB

GLM-4.5

Zhipu AI / Z.ai · Mixture-of-Experts (MoE) · MIT

355B
Total parameters355B (32B active per token)
Memory required426.0 GB

DeepSeek-V3.1

DeepSeek · Mixture-of-Experts (MoE) · MIT

671B
Total parameters671B (37B active per token)
Memory required805.2 GB
Models we deliberately excluded
Models we deliberately left out despite their popularity: Llama (Meta Community License), Qwen2.5 and earlier Qwen releases (custom Qwen license), Command R+ (CC-BY-NC, non-commercial), Falcon-180B (custom TII license), BLOOM-176B (RAIL license), DBRX (Databricks Open Model License), and the original Dec-2024 DeepSeek-V3 (weights under a separate custom license). Each carries licensing terms that are not MIT or Apache-2.0, even though several are excellent, widely-used models.