Joe Brockmeier

RE: https://mastodon.social/@lobsters/116528225865871694

I'm not at all shocked that open-weight models are closing up; a little surprised that the players are showing their hands this quickly, but anybody who has paid the slightest bit of attention to tech over the past 30+ years should have seen this coming.

You cannot depend on big tech to keep things open when there is money to be made in making them (more) proprietary.

It is foolish to plan on anything remaining open when its openness depends on a single vendor maintaining its openness over the long haul.

May 6 at 3:15pmWeb
Show threadAlyssa CoghlanMay 6
@jzb https://apertvs.ai/ is the one example I've seen of a genuinely open model that shares everything about how it was trained while still aiming to operate at a comparable level to the ethical tire fires that are the big proprietary and faux open models (although I'm not clear on the degree to which Mistral are walking the walk when it comes to ethical model training, rather than just talking the talk).
APERTVS.ai

Fully Open Foundation Model for Sovereign AI

Show threadAlyssa CoghlanMay 6
@jzb Hmm, that was confusing phrasing on my part (Apertus isn't from Mistral, but Mistral seem to be the least obviously unethical of the notable proprietary LLM providers, so I was trying to preempt any "What about Mistral?" replies)
Show threadJoe BrockmeierMay 6
@ancoghlan Curious; have you tried / used the apertvs one?
Show threadAlyssa CoghlanMay 6
@jzb I haven't, I just saw the announcements when they finished training the model and when they published it.
Show threadMatěj Cepl 🇪🇺 🇨🇿 🇺🇦May 6

@jzb

Actually, are there any truly open source (including training data) LLM models?

#AI #LLM #OpenSource #FLOSS #InformationWhatToBeFree

Show threadJoe BrockmeierMay 6

@mcepl I don't know of any that actually supply the training data. IBM Granite is supposed to be really open; it has disclosures about its training sets, but just skimming them I see one data set called "Webhose" that is from private non-publicly available datasets obtained from 3rd parties.

@ancoghlan pointed me at Apertvs.ai which is supposed to be a "fully open model" but I haven't dug into it much myself. Might poke at it a bit when I have some time, though.

https://apertvs.ai/

APERTVS.ai

Fully Open Foundation Model for Sovereign AI

Show threadMatěj Cepl 🇪🇺 🇨🇿 🇺🇦May 6

@jzb @ancoghlan

This is what Qwen3 has to say on the topic https://paste.sr.ht/~mcepl/0104cde37b7c289b8b9bf00b93e3c1399a3a80d5

Show threadAlyssa CoghlanMay 6
@mcepl @jzb I couldn't easily find a definitive answer on Qwen3's actual training cut-off date, but it does appear to be prior to the Apertus model release in September 2025.
Show threadPeter O'ShaughnessyMay 6

@mcepl @jzb Yes but there's only about 5 of them validated by the OSI as being compliant with their definition of open source AI:

"Pythia (Eleuther AI), OLMo (AI2), Amber and CrystalCoder (LLM360), and T5 (Google)"

https://opensource.org/ai

Open Source AI - Open Source Initiative

What’s Open Source AI? Following the same idea behind Open Source Software,an Open Source AI is a system made available under terms that grant users the freedoms to: Benefits of Open Source AI...

Open Source Initiative
Show threadKyle RankinMay 6
@jzb The most exciting advances in open weight models are coming from China anyway, and it seems at least for now they are generally staying open. I'd consider the Qwen3.6 series the flagship for models you could run on a personal laptop.
Show threadNeal Gompa (ニール・ゴンパ) May 6
@kyle @jzb Not that this stuff excites me *that* much, but Gemma 4 bucked that trend a little bit too: it is an open weight model, with everything licensed Apache-2.0.