←
Keith18,146 posts
Keith avatar

Keith

@gnukeith

Professional larper @Brave @Obscuravpn Researching @ PhD in sleeping, prev a penguin

◷ Joined May 2023▧ 4,591 Photos and videos
Keith avatar
Keith@gnukeith·1h

I am far more excited for this release than full Hy4, I've been thoroughly impressed by this model, it isn't a coding model but it is an awesome subagent for a bigger model / research buddy, when I put it against Haiku 5.5 I was fully expecting it to lose against it but I found it was on-par, lost in some scenarios, won in more than in lost. If you have been using Qwen 3.8 Flash Next I think this is an upgrade.

Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-f… 🐙: github.com/Tencent-Hunyuan/H… 📄: hy.tencent.ai/research/hy4-f… Released under the Tencent Hy License. See full terms here: huggingface.co/tencent/Hy4-f…

Post attachment
Tencent Hy avatar
⟳ Keith retweeted
Tencent Hy@TencentHunyuan·9h

Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-f… 🐙: github.com/Tencent-Hunyuan/H… 📄: hy.tencent.ai/research/hy4-f… Released under the Tencent Hy License. See full terms here: huggingface.co/tencent/Hy4-f…

Post attachment
Keith avatar
Keith@gnukeith·10h

I knew it after Qwen and Kimi started cracking down on distillation that the American models would hit a wall, sad.

It looks like GPT 6.2 is a benchmaxxed model unfortunately 😭😭

Keith avatar
Keith@gnukeith·10h

Bandwidth is purely a limitation of your mind, it isn't real.

Andy Yen avatar
⟳ Keith retweeted
Andy Yen@andyyen·15h

As promised, we're open sourcing and releasing all of our internal AI tooling as products you can use, starting with Lumo Desktop today. Much more coming in the coming weeks and months, the @ProtonPrivacy team is working hard to deliver faster to the community.

1/5 AI should help you get more done, without compromising your privacy. Today, we're introducing Lumo on Desktop, Connectors, Schedules, and Skills. More ways to work with AI, with the same commitment to keeping your data private. 🧵

Post attachment
Keith avatar
Keith@gnukeith·11h

Currently 83 are running at the same time.

Keith avatar
Keith@gnukeith·14h

I think most major labs have lost the plot of what actual intelligence looks like, I think the whole dream is to have a model with so much knowledge about absolutely every possible scenario ever created and will ever be created but it loses all value of being smart.

Keith avatar
Keith@gnukeith·14h

People have been telling me for months that benchmarks mean something, I've always said choose the model you vibe with rather than how smart it is. I'd rather have a chill buddy that tries it's best rather than a 8x PhD student from top universities talk to me in their own frat guy language.

Keith avatar
Keith@gnukeith·14h

Choosing a model because how well they do in a benchmark is cope, GPT 6.2 Astra / Sol are great example of that. This model is incapable of understanding what matters in a research scenario, it focuses too heavily on doing the task efficiently it ends up taking significant shortcuts. Opus 5.5 is still the best overall, Deepseek V4.1 Flash and Grok is great as a background model. I'm trying it in an orchestrated agent format (yes I had 30 agents running I am doing some research). If you are doing ANYTHING important and you actively hate your job, use the new Sol and Astra models. Inkling 2 Small is a much better model than Sol, it maybe isn't as smart on paper but it cares enough to get things done for you. If you use ChatGPT, get legitimately ANYTHING else, Qwen 3.5 9B, I'll do you one even better. I will answer your questions! That would be more useful than whatever this thing is.

Post attachment

I've been writing a prompt for half an hour to test Deepseek V4.1 Flash, Grok 4.7, Opus 5.5 and GPT 6.2 Astra to see if this new model is any good. I shall report back (this will take couple of hours).

Keith avatar
Keith@gnukeith·17h

FWIW this is challenging because it focuses on inference architecture knowledge and research it needs to be extremely broad with it's searches. Especially on the AMD platform where there isn't a whole lot about the 395/495 models.

Keith avatar
Keith@gnukeith·17h

I've been writing a prompt for half an hour to test Deepseek V4.1 Flash, Grok 4.7, Opus 5.5 and GPT 6.2 Astra to see if this new model is any good. I shall report back (this will take couple of hours).

Obscura: The Privacy-first VPN avatar
⟳ Keith retweeted
Obscura: The Privacy-first VPN@obscuravpn·20h

Many VPNs Are Worse Than No VPN. Popular VPNs, especially free ones, log, leak, or spy on the people they promise to protect. Here’s what to watch out for, and why Obscura is different.

Keith avatar
Keith@gnukeith·20h

Unless they become really cheap, that's my clause!