Is that "open-source" AI actually open? A 4-level framework, three models scored
alex_verem · x · 2026-10-21
A practical guide to judging how open an "open" AI release really is: read the license, not the launch post. The piece lays out a 4-level framework based on the OSI Open Source AI Definition — open access (API only), open weights, open weights + open license (Apache 2.0/MIT), and fully open (weights, training code, data).
Key event: on Oct 15, 2026, Comfy Org launched a $25,000 contest for open rebuilds of Higgsfield's creative apps, saying Higgsfield promised to open source its platform but never shipped the code and workflow files.
Radboud University researchers track this in the European Open Source AI Index, scoring systems on 14 openness dimensions. Applied to three new models: DeepSeek-V4.1-Flash (MIT license, weights open, training code closed, server-only local runs) and MiniCPM-V 4.7 (Apache 2.0, runs locally) both land at "open weights"; Ultramarine 1 Flash has no open weights at all, making it open access for now. Most releases called open give you less than the word promises.
More from Models
- NVIDIA releases seven open model families at GTC Berlin, from Nemotron 3.5 to Cosmos 3.5 and BioNeMo — NVIDIAAI · 2026-10-21
- NVIDIA Expands Open Model Family: Nemotron 3.5, Cosmos 3.5, Kumo, BioNeMo and Ising — NVIDIA Blog · 2026-10-21
- Humansand team members tout Ampersand launch: 'towards human flourishing with superintelligence' — niloofar_mire · 2026-10-21
- Jensen Huang says NVIDIA leads in shipped open models, with Alibaba and HuggingFace next — testingcatalog · 2026-10-21
- Google reportedly prepping new Gemini models Carbon and Argon weeks after Gemini 4 — haider1 · 2026-10-21
- Qwen3.8-Flash-Next vs Opus 5.5: 12-Run Rust Benchmark Shows 87% vs 68% at 1/50th the Cost — deepu105 · 2026-10-21