Harshit Pandit

My new blog is out now!!! Apple Silicon is seriously underrated in running large language models. I literally got upto 200 tok/s on my M5 !! 🔥🚀 Read to find out more: harshit147.dev/blog/qwen-on-a…

@nvidia do you guys have an extra 5090 lying around? 👉🏻👈🏻

I think Companies like Nvidia, Apple, Modal, NASA, Autodesk, IBM, Cisco, Qualcomm have some real engineers and they make actual 'Computing' products that somewhat closely match to my degree. They provide the tools and tech for someone to practice and learn computer engineering.

I'm so sick of being a coding kitty and using all the fancy shit like Claude Code. I have an engineering degree FFS, I studied calculus, signals, electronics, graphics. I want something real, deep and be an actual 'engineer' and understand, profile and build real shit.

"We're more afraid of people instead of the technology getting better" - An absolute W take by @neetcode1 Check out his latest video: youtu.be/wUdjDWFz7RQ?si…

Most people don't even need a trillion parameter model. A simple decent model running on their own hardware is more than enough literally.

Does anybody know what's the best way I can learn and practice CUDA for free? Or anything similar for Apple silicon?

Cursor started out as 'just another VS code fork' but it seems that they're moving more towards building a real product. No wonder SpaceX acquired them for $60B

@mariodian @HuggingModels Yeah and it works wonders! I've only got a base M5 24GB RAM so my mac heated up a lil bit but the quality was top notch.

@Harshitp147 @HuggingModels Have you used them with OpenCode? I’m starting to think it may be related. I tried on another more powerful mac, and I’m running into the same issue of reasoning being cut off mid-execution with no apparent error.

Gemma 4 12B Coder is here and it's a game changer for local code generation. This GGUF model packs Google's latest gemma-4 architecture into a compact 12B size, perfect for running on consumer hardware. It's optimized for reasoning and thinking, making it ideal for developers who want fast, private coding assistance without the cloud.

@mariodian @HuggingModels It's probably the model itself then. I've used original Qwen, Gemma and Liquid and have rarely encountered such issues.

@Harshitp147 @HuggingModels It's equally unusable via llama.cpp. It says it will do something and then it just stops reasoning mid-execution. I don't think it can use tools.

@mariodian @HuggingModels Then I think you're better off with llama.cpp with MPS, still better than LM studio imo

@Harshitp147 @HuggingModels Unfortunately, there are no weights available, only quantized models, so I can't convert it. huggingface.co/yuxinlu1/gemma…

@DRexroat4841 Is that from Circle K?

Happy Monday to everyone!

@mariodian @HuggingModels Yeah u convert the model, but luckily the package comes with both conversion as well as the server scripts, so a few commands and you're all set!

@Harshitp147 @HuggingModels You'd have to convert the model, right? Also, I don't think mlx-lm exposes an API endpoint so that I can use it in OpenCode.

@mariodian @HuggingModels Try to use mlx-lm instead of lm studio. Best optimization and performance for Mac

@HuggingModels I’m testing it on an M2 Pro Mac with 32GB RAM (will test with an M4 Pro 48GB RAM later) using OpenCode. It gives short answers and doesn’t follow instructions well. It basically dies mid-execution, but the backend seems fine. Any specific settings for LM Studio?

Coffee + Claude Code + Clouds = Vibe ☘️✨🚀

Something was happening today at Dames street 👀

@K_aditya25 Amazing blog! Well written 🙌🏻

Getting a job or an internship in Ireland as a software student/grad is challenging. I have lost count of the number of applications I've made. I wrote this article summarising the truths and lessons I wish I knew 3 years ago when I started looking for internships in Ireland as a Computer Engineering student. The lessons shared in this article don't just apply to non EU/UK students in Ireland. This should benefit all students and graduates. Hope this helps :)

You don't know inferencing and that's why your local model sucks on your MacBook. I dove deeper about local inferencing in two engines, and the results were shocking!! Check out harshit147.dev/blog/qwen-on-a… We inferencemaxxing our models with this one🚀

DO NOT ASK YOUR CODING AGENT TO GIT COMMIT FOR YOU. Go to the terminal and do it yourself. If you do so, it will write a nice commit message and add a "Co-Authored-By", which gives them the right to use your codebase as their training data.

@Brennan_Lup Pilots go through rigourous training and learn how to fly manually before they're 'vibe-fly' the plane with autopilot.

Be honest devs, is coding still worth learning in the AI era?