Hacker News

points by notenchange 3 days ago | 0 comments

I’m trying to do some GPU unlocking work with some oldish mining cards (NVIDIA CMP 170HX 10G) to fully unlock them to be used in LLMs, and Sol 6.1 keeps refusing to do the work on the basis of “cyber concerns”. It’s highly frustrating. Sonnet 5.5 is even worst.

Luna 6.1 is fine with it, but the results aren’t good enough. I’m going to try Terra 6.1 (from the post, I should expect about the same results). And hope it doesn’t complain as much about these non existent cyber security risks that for some reason it keeps raising.

P.S. I also tried Deepseek Flash 4.1, but the results simply didn’t cut it.

P.P.S. This all investigation was a major eye opener on why we need to run local models or be forever tied to some artificial limits to AI for the masses.

[–] micw 3 days ago

Try GLM for such things. Better than DS Flash in my observations but less restricted than GPT

[–] Sammi 2 days ago

New hn post on deekseek and GLM models that are unrestricted.

https://news.ycombinator.com/item?id=50254836

[–] Duwensatzaj 2 days ago

Abliterated models have always struck me as severely brain damaged. Maybe that’s gotten better.

[–] slopnt 2 days ago

that has actually been proved to be the case. the weight tweaking being done by abliteration does change token distribution in some cases precisely making LLMs produce tokens it wouldn't have otherwise (and therefore less reflective of training data)

[–] krzyk 2 days ago

Nitpicking, but I had to chekc if they introduced it - there is no Luna 6.1, there is Luna 6.

[–] jeffy29 2 days ago

There is no Luna 6.1, yet.

[–] throw20261022 2 days ago

Their model release pattern has been a joke. Any firm in their right mind would have released the large+mid+small tier models at the same time. OpenAI though can't seem to pull it off. It's like they either don't have the hardware for it, or they have to wait to collect substantial user data over which to distill the tiers. Either way it's disruptive to user workflows, and not sane to invest in.

[–] senordevnyc 2 days ago

This seems like a really trivial thing to get so worked up over. Workflows work fine with a mix of different model versions. (I do it all the time, including open weight models, different reasoning levels, etc. Sounds like you need some evals.

Also, you can just…wait a little longer until all the models for the version number you want are available? You don’t lose anything, so how is not “sane” to do so?

[–] basch 2 days ago

staggering has to be a better use of resources, and I would think most workflows can accommodate using a 6.0 of one model and a 6.1 of a completely different one without balking.

[–] throw20261022 2 days ago

> most workflows can accommodate using a 6.0 of one model

No, they can't, considering Terra never existed for 6.0. It's a mess, and no self-respecting user would care for it.

[–] aand16 2 days ago

Kimi K3.1 is out