Bart Slodyczka

In this video I show you how to use DeepSeek Harness completely free by routing requests through OmniRoute instead of paid Claude/ChatGPT APIs or running local AI. You get ~1.5B tokens/month (and ~2B in the first month with sign-up bonuses). 🏆 Support the channel: https://www.youtube.com/channel/UC8F7X90NDmMr1mGx8897r_A/join 👉 One-on-one coaching, team coaching, or agent builds: https://workingmodels.ai/start 👉 Watch all my Local AI videos: https://www.youtube.com/playlist?list=PLi7jtY2ZZqRaidYBIucDnKnhD7Jmt-2sY 👉 Connect with me: https://x.com/bartslodyczka

Learn how to build beautiful websites using local AI models. Try Higgsfield here: https://higgsfield.ai/s/higgsfield-general-campaign-bartslodyczka-DWgNWR Full video: https://youtu.be/0Yin3Er9DX0

I tested Ornith-1.5 35B MoE at 4-bit on MLX with one-shot prompts: invoice vision, a tower defense game, a spreadsheet with live formulas, and a Three.js FPS. Vision hit 98%, the games and spreadsheet mostly worked, and Three.js broke every time on a single missing character. All prompts and results are in the GitHub repo linked in the comments. 👉 All tests in my Github repo: https://github.com/Barty-Bart/ornith-1.5-35b-tests.git 👉 One-on-one coaching, team coaching, or agent builds: https://workingmodels.ai/start 👉 Learn how to build 3D animated websites with Qwen 3.8: https://youtu.be/0Yin3Er9DX0 👉 Try Higgsfield here: https://higgsfield.ai/s/higgsfield-general-campaign-bartslodyczka-DWgNWR 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Watch all my Local AI videos: https://www.youtube.com/playlist?list=PLi7jtY2ZZqRaidYBIucDnKnhD7Jmt-2sY 👉 Connect with me: https://x.com/bartslodyczka

In this video I show you how to build animated websites using Qwen 3.8 27b, pi.dev, and Higgsfield AI. Try Higgsfield here: https://higgsfield.ai/s/higgsfield-general-campaign-bartslodyczka-DWgNWR This is my take on the popular 'Build $10k Website With Claude Code' video format, where we use local AI models to do the work. I built this site using Qwen 3.8 27b 4-bit MLX and Higgsfield AI. We used Higgsfield to generate our images and videos. The video was needed so we could add the scroll-scrub effect within the website, where as you scroll, you flick through the frames, and give the 3d visual. Very surprised with the 4-bit quant on this build as I didn't run into any issues. No looping, no errors with code generation. And in general, Qwen was smart, capable, and proactive. I also ran this same process with DeepSeek V4 Flash 0731 and it was just as amazing. Local AI models are getting really good. 👉 One-on-one coaching, team coaching, or agent builds: https://workingmodels.ai/start 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Watch all my Local AI videos: https://www.youtube.com/playlist?list=PLi7jtY2ZZqRaidYBIucDnKnhD7Jmt-2sY 👉 Connect with me: https://x.com/bartslodyczka Timestampes 00:00 Website Preview 00:30 Quant & Tools 00:48 Connecting Qwen to Pi.dev 01:45 Setting Up Higgsfield MCP 02:42 Generate Reference Image 06:20 Website Part 1 08:23 Generate Scroll Scrubbing Video 10:28 Embed Video Onto Website 12:52 Website Edits 03:44 Anthropic Front End Design Skill

Qwen 3.8 27B just dropped, so I ran it against DeepSeek V4 Flash 0731, a model ten times its size, with both downloaded and running locally on my Mac Studio. I gave them three coding tasks of increasing difficulty: a weather dashboard pulling live data, a tower defense game, and an Excel style spreadsheet with working formulas. The 27B model held its own and on the hardest test it was the only one that produced something that actually worked. 👉 Github Repo for all tests, results, machine specs, and model performance specs: https://github.com/Barty-Bart/qwen-vs-deepseek-local-coding 👉 One-on-one coaching, team coaching, or agent builds: https://workingmodels.ai/start 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Watch all my Local AI videos: https://www.youtube.com/playlist?list=PLi7jtY2ZZqRaidYBIucDnKnhD7Jmt-2sY 👉 Connect with me: https://x.com/bartslodyczka

In this video I run DeepSeek V4 Flash 0731 locally on my Mac Studio M3 Ultra and test whether it can do real work. It scores 52 on the Frontier Language Model Intelligence Index, which puts it level with GPT 5.4 and ahead of Claude Sonnet 4.6, so I plug it into the Pi agent harness (pi.dev) and give it three tasks I do every week: diagnose and fix a broken n8n workflow and make it production ready, build a multi part inventory reorder report from Excel, and audit a ClickUp workspace to find repe

In this video I give you an overview of LM Studio Bionic. Bionic is a free local AI app that replaces Claude Code and Claude Cowork with built-in Work Mode, Code Mode, real-time voice transcription, MCP integrations, and local web search. You can run both local AI models you download on your computer, and other open models like Kimi K3, GLM 5.2 and more. This is a super exciting new local AI app! Enjoy :) 👉 Get in touch: [email protected] 🤔 Do you want a course on this topic? https://www.workingmodels.ai/survey/ 📬 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Connect with me: https://x.com/bartslodyczka Watch Next: Local AI Playlist https://www.youtube.com/playlist?list=PLi7jtY2ZZqRZUBEsWt6kyHqwtQV2FplGj

In this video I build an AI employee that runs my customer support inbox using Pi (pi.dev), the open source agent harness that powers OpenClaw. It's powered by a local AI model running through LM Studio (Qwen 3.6 35B), reads new Zendesk tickets, and sends customer responses. Then I set up a cron job with Pi's print mode so it works 24/7 on its own, and the same setup works for Gmail or Outlook, fully local and private with no API costs. 🤔 Do you want a course on this topic? https://www.working

In this video I test out the new open source agentic coding model Ornith 1.0 9b on my 16gb M4 Mac Mini. We load it up in Lm Studio and understand the full memory requirement for weights + context, we plug it into pi agent harness to run it through a coding task, we then compare results to that of the bigger Ornith 35b model, and finally we understand if the 9b is worth running locally for coding tasks. 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.

Head to my channel for more local AI vids Or read my blog for my ai experiments: https://blog.workingmodels.ai/

Read more: https://blog.workingmodels.ai/p/glm-5-2-local-mac-studio

In this video I use Claude Code to set up Hermes Agent on a 16GB Mac Mini and plug it into a local Gemma 4 12B model, so the whole thing runs 24/7 with no API bills. I let Claude do everything from the terminal: changing the Mac's settings to stay always-on, installing and wiring up the agent, and connecting it to 3rd Party Tools so the agent actually does real work. 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Connect with me: https://x.c

Fable 5 is leaving the Claude Sub on June 22. We have 10 days left. Stop building flashy things and start building stuff you will actually use. Here are 5 things I am doing. If you are financial able to, you should consider trying the same.

Testing the new Gemma 4 12b on my 16gb M4 Mac Mini. This is a free AI model you can download and run on your computer. Can get around 40k context window. Follow for more Gemma 4 videos!

Google just dropped Gemma 4 12B, a free open-source model small enough to run on a 16GB Mac Mini, and in this video I give you an honest breakdown. We use LM Studio to run the model and VS Code to preview the HTML app we build. 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ 👉 Watch all my Local AI videos: https://www.youtube.com/playlist?list=PLi7jtY2ZZqRZUBEsWt6kyHqwtQV2FplGj 👉 Connect with me: https://x.com/bartslodyczka 👉 Get in touch: b

In this video I run Claude Code's new Dynamic Workflows feature to spin up 1000 AI agents on one task, completely free, using local models in LM Studio instead of the paid Anthropic API. I walk through the full setup start to finish: the Claude desktop Gateway, loading a local Gemma model, the alias trick that makes Claude find it, adding web search with an MCP, and firing a /deep-research workflow that runs hundreds of autonomous agents on your own computer. Works with Ollama and cloud models l

Most people use Claude Cowork at a fraction of what it can do because they treat it like a chatbot. You ask, it answers, you still do the work. The shift happens when you stop running sessions and start running projects, with connectors, context, and scheduled tasks wired in so Cowork carries the work for you. In this video I walk through how to think differently and set up your first real project. 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/

Local AI is getting insane. This video shows you exactly how to run free, private AI models on your own computer using LM Studio — no internet needed, no monthly fees, your data never leaves your device. What you'll learn: - The difference between cloud AI (ChatGPT, Claude) and local AI - How to pick a model your hardware can actually run - Setting up LM Studio and downloading your first model - Adding MCP tools to your local AI: • Brave Search — web search inside the chat • Playwright —

In this video I show you how to run Claude Cowork 100% free and 100% private using Ollama and local AI models that run on your own computer. This is the new Cowork on 3P. I also show how to connect OpenRouter to Claude Cowork to access hundreds of other cloud AI models. 📬👀 Subscribe to my blog for more tests, tips, and walkthroughs: https://blog.workingmodels.ai/ ** THIS METHOD IS NOW OUTDATED - WATCH THIS VIDEO TO MAKE IT WORK**: try this: https://youtu.be/ah_Bl_aGQF0 👉 Watch all my Claud

In this video, I break down how to use Claude Cowork as a virtual assistant that can automate 99% of your tasks — from installing the desktop app and choosing the right plan, to understanding the critical difference between Claude Chat, Cowork, and Code. I then walk you through my exact workflow for creating projects, loading them with context so Claude actually understands your world, and setting up scheduled tasks that run autonomously (like tracking flight prices daily on Skyscanner). By the

In this video, I break down everything you need to know to use Claude Cowork like a pro — from setting it up for the first time to building a structur...