Revision history
- correction
Added a section on language-optimised models, which a capability ranking structurally hides.
- Added Rakuten AI 3.0 — a ~700B MoE under plain Apache 2.0, developed under Japan’s GENIAC programme. It belongs in any discussion of Apache-licensed models at frontier scale and was simply missed.
- Added Preferred Networks’ PLaMo 3.0 Prime and LLM-jp as pointers, and noted the same blind spot applies to Indic, Arabic and Southeast Asian models.
- revised
Reframed for professional multi-GPU deployment after the first version read as a home-hardware guide, and fixed an ambiguous column that conflated two different numbers.
- Split “memory to serve” into two columns — “Download (disk)” and “GPU memory to serve”. They are different numbers, and conflating them is how capacity planning goes wrong: Qwen3.8 is 4.89 TB on disk at BF16 but 2.5 TB of HBM if you serve the FP8 checkpoint.
- Scoped the article to 2-GPU nodes up to NVL72-class racks; dropped the single-GPU/home tier.
- Replaced the ten per-model prose sections with a “Best for” table column, and added a hardware-tier table mapping available HBM to the models that fit.
- Added a “Licence and capability gotchas” section so the findings buried in the deleted prose survived — Qwen3.8’s open weights being text-only, GLM’s MIT→bespoke regression between 5.2 and 5.3, Kimi’s licence change.
- Added download links for all ten models, each verified to resolve.
- Removed an unverifiable third-party claim about a 600B Meta model that appears on no first-party channel.
- first published
Initial survey — the ten most capable freely downloadable models and five you pay to self-host, with parameter counts, checkpoint sizes and licences checked against primary sources.
Revision tracking was added to this site on 2026-10-05. Entries dated before then were reconstructed from the repository's commit history and deployment logs, so they record what changed but not necessarily every intermediate edit.