David Dalcu

I build things, local LLM engines & UI’s. Created for Desktop and MLX Chat for iOS. (Zig based engine)

[David Dalcu @ddalcu]

Free macOS menu-bar app for local AI: chat, agent mode, Control+Space quick launcher, image/video gen, voice cloning.

All on Apple Silicon. No cloud.

https://t.co/eVMeE12nh4

[David Dalcu @ddalcu]

@ivanfioravanti @_tdubey Laguna's attention is dense bf16 (only the experts are nvfp4)

[David Dalcu @ddalcu]

@ivanfioravanti @poolsideai @ollama How does it compare to DS4F and HY3 ? It’s all about that intelligence per GB factor.

[David Dalcu @ddalcu]

@ivanfioravanti Nice to see big tech participating in the open model space. Would be nice to see a good and efficient video model.

[David Dalcu @ddalcu]

@jack What happened to bitchat?

[David Dalcu @ddalcu]

@lexrus Yep, 100% agree.

[David Dalcu @ddalcu]

I would not be surprised if Apple decides to get back into the server space and build Mac Pro AI Racks... people are already building racks with the Mini's & Studio's. Why not build something specific ? I would not like it.. but it makes perfect business sense.

[David Dalcu @ddalcu]

It really depends on the engineer. I put a lot of “weight” on passion, and love for the craft when hiring folks.
Especially today. 10x engineers are now 100x engineers. And 1x engineers, mostly fall in the vibe coding category. The middle doesn’t exist much, it’s like inverted bell chart.

[David Dalcu @ddalcu]

@antirez wow, nice !

[David Dalcu @ddalcu]

@0xGramajo Neither, one is slow, and one doesn’t have enough vram. Mac Studio !

[David Dalcu @ddalcu]

@Kevrsub Have you seen the amazing design of the Mac Mini power button ? 🤣 ….they did a terrible job on both.

[David Dalcu @ddalcu]

@ivanfioravanti Haha… I’m “game” ! I got a few ideas and been researching a bit towards this direction, but first step I need a fast way to run local models 😅

[David Dalcu @ddalcu]

Upcoming, run @pidotdev or @NousResearch Hermes in a sandbox, pointed to a fast local MLX model ! No docker or system dependencies required, or complicated setup process. Only a few clicks.

[AICodeKing @aicodeking]

Qwen 3.8 Max is actually a very good model. Outperforms all Opus models in my benchmarks. There are some issues in tool calling with some harnesses. But, the raw intelligence is just crazy.

[David Dalcu @ddalcu]

@christianelton_ This is a very interesting topic to me, but rather than making it $$$ based, I would like to make it Peer2Peer & and "Internet points" / Leaderboard based. I have already built a proof of concept done for this in https://t.co/2m7ZVWa52X and working on adding it to MLX-Serve

[Qwen @Alibaba_Qwen]

Qwen3.8 is launching and going open-weight soon!🌐

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

Can't wait to hear what you build. Stay tuned! 🚀 

Token Plan
international：https://t.co/YRvcGdB9Bv
China：https://t.co/PKMUNwUuRp

[Ivan Fioravanti ᯅ @ivanfioravanti]

Are you ready for a new mega release of DwarfStar next week? 🚀

[Jun Song @jun_song]

Best Local AI models by VRAM size (7/18) :

Phone (~4gb) : Bonsai-27b
Nano (~12gb) : Gemma4-12b
Micro (~36gb) : Qwen3.6-27b
Small (~128gb) : Deepseek V4 Flash ds4
Medium (~256gb) : GLM-5.2-2bit
Large (~512gb) : GLM-5.2-NVFP4

[David Dalcu @ddalcu]

@Kevrsub Correct, I could not find anything either. I have all the correct ingredients to build support for this in MLX-Serve, so... let's see what happens !

[David Dalcu @ddalcu]

Would it be a good idea to have a native Apple Silicon auto GGUF->MLX translation layer at load-time ?

Or not worth it because MLX Models are plentiful these days..