←
Dr. Daniel Bender9,070 posts
Dr. Daniel Bender avatar

Dr. Daniel Bender

@drdanielbender

I'm building my assistant in Hermes. No, I won't blindly trust it ◆ I teach you the way of responsible AI: your data, your rules ◆ PhD in computer science ◆ Dad

◷ Joined January 2022▧ 1,653 Photos and videos
Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·11 Jun 2025

Ever wondered how model quantization (FP16, Q8, Q4) *really* affects performance? There's an analogy that makes the trade-offs crystal clear... and it involves something you might drink. 😉🍺 Kudos to @jtdavies for this brilliant comparison. 🙏 See the image for the full explanation! 👇

Post attachment
Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 17

The space collapsed after seven minutes, with so far no one who had joined. I am not sure if people tried to join but were unable to, but the result is the same: no space today, and we meet again in two weeks at the same time. 😊

Mia avatar
⟳ Dr. Daniel Bender retweeted
Mia@MiaAI_lab·Nov 16

The most important part of this post is that they're releasing open weights for multiple smaller Qwen4 models. That likely means Qwen4-27B and Qwen4-Flash-Next. Qwhen though?

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

Post attachment
Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

Open-weight models are well established now. The gap to the top of this chart looks large, but all four are capable of plenty of useful work. I use today's closed frontier models for the hardest tasks. Most work can now be handled with open-weight models in more private setups.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

That's my current roundup. I'm especially looking forward to Qwen4 open-weight releases that reshuffle it. Flash-Next already previews its architecture. @Alibaba_Qwen teased the new Qwen 4 release earlier this month.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

Qwen3.8-27B is the strongest small-model pick right now: 20 points at just 27B parameters. With Q4 quantization, it fits on a 24GB GPU like NVIDIA's RTX 3090 or a Mac with 32GB system memory. Context length still matters, but this is a genuinely practical local model.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

Qwen3.8-Flash-Next scores 29 with 180B total / 6B active parameters. You can run it with consumer gaming GPUs and 64GB+ system RAM using the right quantization and offloading setup via Strata by @coldniko or on systems with 128GB shared memory like the DGX Sparks or Strix Halo.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

DeepSeek V4.1 Flash is my API sweet spot: 33 points, 552B total / 16B active parameters. The extremely high @OpenRouter usage makes sense: capable enough for a lot of real work, with very competitive token pricing. This is where intelligence meets a usable budget.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

Kimi K3.1 Pro is the clear intelligence leader among the open-weight models in @ArtificialAnlys intelligence index: 46 points. At 2.8T total / 104B active parameters, it's also the biggest model here. Of my four picks, it's the expensive option for both hosting and API use.

Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 13

The most interesting AI model isn't always the smartest one. Sometimes it's the one that runs compliant in the region you live in or on a GPU you already own. My current open-weight shortlist: 4 models, from Kimi's 2.8T parameters down to Qwen's 27B. Details about them in the 🧵

Post attachment
Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 9

A source link doesn’t guarantee an AI answer is right. I asked ChatGPT and Hermes Agent a simple question: What's a soccer club’s annual membership fee? They gave different answers. Both cited sources, but only one was correct. Check the origin of the sources yourself!

6:06
Dr. Daniel Bender avatar
Dr. Daniel Bender@drdanielbender·Nov 5

WTF and thanks to OpenAI for making my digital life a bit safer by spending an equivalent of $1.2M to secure the Linux kernel🙏 Yes, less costs than the stated $1.2M for OpenAI as they don't need to pay the margin of customers, BUT compute is a rare resource for everyone. 👏

Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline. In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.

Post attachment
Simon Høiberg avatar
⟳ Dr. Daniel Bender retweeted
Simon Høiberg@SimonHoiberg·Nov 4

I'm so happy I'm self-hosting everything now! If you're not, please stick with the big providers: AWS, Google Cloud, Azure. I wouldn't trust any of these small PaaS - Railway, Supabase, Netlify, Vercel, etc - for ONE second. Get off the cloud. (Or stick with the really big ones).

Bro, you can't fucking say openly on social media you just attempted to compromise an organizations infrastructure under your employer and full legal name Are you actually fucking insane?? Genuinely what the fuck are you thinking??? And how the fuck does this dude know the employees name? Why did the employee openly admit "hehe ye thats me, but no mallace intended, it was my AI agent" ????? WHAT. THATS STILL A CRIME. WHAT ARE YOU TALKING ABOUT.

Post attachment