AI News Today | Julian Goldie Podcast

Qwable 27B Coder: Best Local Coding Model Yet? (Benchmarks, Demos & How to Run on Mac)The video demos Qwable 27B Coder, a new open-source local coding model on Hugging Face (updated end of June) built on a Qwen 3.6 27B base, showing projects made locally like a polished animated landing page and working games. It’s run on a Mac Studio (Apple M4 Max, 36GB) using Apple MLX (not available via Ollama) and is integrated into an agent operating system to generate live previews and save builds in a workspace. The presenter says Qwable tops their local leaderboards (Goldie Bench/Cody Bench comparisons) and outperforms recently tested local models like Gemma 4 12B Coder, Quifos 9B, and Onif 1.0 on the same tasks, though it runs noticeably slower than smaller models such as Quifos 9B.00:00 Meet Qwable 27B00:56 Demos Landing Pages01:23 Model Specs Setup02:04 Benchmarks Versus Locals02:40 Running On Mac03:03 Agent OS Live Previews04:24 Speed Tradeoffs05:00 Real Task Comparisons05:49 Local AI Is Accelerating06:05 Agent OS Features Tour06:40 Community Courses Pitch07:28 Wrap Up Thanks

Agent OS Q&A: Using Codex, Claude Code, Local LLMs (Ollama), GLM 5.2, Updates, Memory & Team SetupJulian answers recent community questions about his Agent Operating System, a system for linking and automating AI agents with tools like voice control (Hermes Jarvis), Memory Galaxy, Kanban orchestration, and a local agent engine. He explains multiple ways to use a Codex subscription inside Agent OS (Hermes, image generation via Codex OAuth, the Codex tab, and a new Codex plugin in Claude Code), emphasizes focusing on repeatable systems over prompts/models, and shares thoughts on Qwen 3.6 vs Qwen Agent World. He covers connecting Codex/Claude Code to local models via configuration or Ollama, highlights GLM 5.2’s strong performance and open weights, and advises on simplifying Agent OS for teams with hidden UI sections and SOPs. He also suggests Agent OS video automation/Remotion for SaaS promo videos, recommends GLM 5.2 for Ollama-first setups, explains how to update Agent OS via the latest zip/update file, and describes using an Obsidian vault as shared memory for individuals and teams.00:00 Agent OS Overview00:48 Using Codex Inside02:24 Systems Over Prompts04:01 Qwen Agent World04:37 Local LLM Connections05:16 GLM 5.2 Benchmarks07:08 Sharing With Teams09:14 AI Video Creation10:37 Best Ollama Model11:48 Updating Agent OS12:27 Memory Setup Choices13:06 Team Shared Brain14:24 Wrap Up And Invite

4 Free Ways to Use Hermes Agent (News Portal, OpenRouter, Local Models, OAuth)The script explains four ways to use Hermes Agent for free and ranks the options, highlighting how to switch models and manage setups. It covers using free models that appear on News Portal, connecting Hermes to OpenRouter with a free API key and selecting free models (with a preference for models like Gemma 4 or Owl Alpha), and running local models via Ollama or LM Studio (including pulling models like Ornif and choosing hardware-appropriate options). It also shows how to create separate Hermes profiles and use “Blank Slate” to reduce tool usage and token consumption. The fourth method is OAuth, letting existing subscriptions (e.g., ChatGPT/OpenAI or Groq via Twitter) work in Hermes without extra cost. The video argues this approach improves privacy, flexibility, and cost control, and it promotes the creator’s AI Profit Boarding community and Agent OS setup.00:00 Free Models Overview00:18 News Portal Picks01:09 OpenRouter Free APIs02:42 Profiles and Agent OS03:35 Run Local Models05:09 Create Lean Profiles06:34 OAuth With Subscriptions07:38 Cut Tokens and Stay Private08:43 Agent OS Features Tour09:45 Community and Wrap Up

Hermes Mixture of Agents: Build a Model Council That Beats Single Frontier Models (Goldy Bench #2)Hermes released Hermes Mixture of Agents, which the speaker integrates as a “Hermes Council Engine” tab inside their Agent OS to run a panel of frontier models (e.g., Opus 4.8 and GPT 5.5) that answer privately while a third “chair” model judges and fuses a better final response. They tested it across a 42-task leaderboard on Goldy Bench where Hermes Mixture of Agents ranks #2 overall, behind Fusion, and they show side-by-side examples versus Fusion and Claude Opus 4.8, claiming Hermes outputs are often less buggy and more usable. The episode argues to “stop chasing the model” and instead build swappable systems that squeeze more from existing models, and it promotes the Agent OS/AI Profit Boardroom bundle with the Mixture tab, Fusion and Sakana panels, shared memory dashboard, tutorials, coaching, and setup roadmap.00:00 Hermes Mixture Intro00:41 Council Engine Explained03:53 Leaderboard Results04:38 Showcase Game Outputs05:44 Fusion Comparison06:32 Opus Side By Side09:20 Stop Chasing Models10:08 Agent OS Workflow Demo11:36 Get The Agent OS12:06 Why Systems Win14:34 Recap Key Lessons15:41 Join The Community16:31 Final Thanks

Ornif 1.0: Free Self-Learning Local Coding Model (Runs on Ollama + Works with Hermes Agents)The video tests Ornif 1.0, a new free self-learning local open coding model from Deep Reinforce, showing benchmark comparisons where it outperforms models like Gemma 4 31B and even compares well against larger flagship options. The model can be downloaded via Ollama with a simple command and is integrated into the creator’s agent operating system, including Hermes profiles for agentic workflows. Ornif’s key idea is “self-scaffolding”: it writes its own plans/instructions, generates multiple solution rollouts, gets graded, and improves both planning and coding through reinforcement learning. The creator highlights a gallery of apps, games, tools, and visual outputs made with Ornif, plus local and frontier leaderboards, and demonstrates local offline systems like a Hermes engine, agent Kanban orchestration, and local chat, emphasizing local AI benefits like being free, private, fast, and improving.00:00 Meet Ornif 1.000:34 Setup With Ollama00:45 What Ornif Is01:04 Real Build Examples01:51 Benchmark Reality Check02:11 Self Improving Framework03:02 Reinforcement Learning Loop03:52 Self Scaffolding Explained05:31 Local Leaderboards Demos06:21 Agent OS Integrations07:32 Model Comparisons Use Cases08:05 Join The Community09:05 Final Thanks

Hermes Mixture of Agents: Build a Model Council That Beats Single Frontier Models (Goldy Bench #2)Hermes released Hermes Mixture of Agents, which the speaker integrates as a “Hermes Council Engine” tab inside their Agent OS to run a panel of frontier models (e.g., Opus 4.8 and GPT 5.5) that answer privately while a third “chair” model judges and fuses a better final response. They tested it across a 42-task leaderboard on Goldy Bench where Hermes Mixture of Agents ranks #2 overall, behind Fusion, and they show side-by-side examples versus Fusion and Claude Opus 4.8, claiming Hermes outputs are often less buggy and more usable. The episode argues to “stop chasing the model” and instead build swappable systems that squeeze more from existing models, and it promotes the Agent OS/AI Profit Boardroom bundle with the Mixture tab, Fusion and Sakana panels, shared memory dashboard, tutorials, coaching, and setup roadmap.00:00 Hermes Mixture Intro00:41 Council Engine Explained03:53 Leaderboard Results04:38 Showcase Game Outputs05:44 Fusion Comparison06:32 Opus Side By Side09:20 Stop Chasing Models10:08 Agent OS Workflow Demo11:36 Get The Agent OS12:06 Why Systems Win14:34 Recap Key Lessons15:41 Join The Community16:31 Final Thanks

Qwythos 9B on Ollama: A Free, Private Claude-Style Local Model (1M Context?)The video tests Qwythos 9B, a Claude-style creative reasoning model built on a Qwen 3.5 9B base and available on Ollama, showing how to install and run it locally on a Mac for free with no cloud or token costs. The creator wires it into their Agent OS local engine and demonstrates it building several small apps (to-do list, digital clock, Snake game, landing page, calculator), noting it can produce surprisingly solid designs but can also be glitchy or messy. While it advertises a one million token context window, the practical window depends on available RAM and Ollama settings, and it may cut off or load smaller than expected. They compare it with other local models like Qwikboss and Ornif 9B (with Ornif 1.0 seemingly outperforming) and promote the AI Profit Boardroom and Agent OS for running local models and agents.00:00 Meet Qwythos 9B00:26 Install via Ollama00:43 Agent OS Integration01:39 Apps Built Locally02:29 Speed and Comparisons03:26 Why It Works04:20 Million Token Reality04:54 Model Sizes and Quantization05:36 Pros and Cons06:30 Offline Wrap Up07:09 Plug Into Any Agent07:40 AI Profit Boardroom Pitch08:07 Agent OS Feature Tour08:57 Community and Coaching09:24 Final Thanks

Agent Operating System Q&A: Model Restrictions, Local AI, Hermes Jarvis, SEO & Video AutomationThe episode answers community questions about the Agent Operating System and how to get the most from it, highlighting Hermes Oracle for news and SEO automation, Hermes Jarvis as a voice-activated agent, and managing all agents and client instances from one interface using separate profiles. It argues that gated frontier models like GPT 5.6 previews and the removal of Fable 5 matter less than building robust systems that can swap models in and out, including using alternatives like Fusion or Sakana Fugu and open-weight local options such as GLM 5.2. The host discusses a high-volume posting strategy with strict quality control, rising demand for local model setups for privacy and cost, automating video production via a video agent and avatar workflow training, using Paperclip for fact-checking and multi-agent teams, Windows support, SEO article generation and deployment, UI improvements to the Kanban board, and examples of AI-assisted game creation, while promoting the AI Profit Boardroom for daily Q&A, tutorials, coaching calls, and the Agent OS zip file.00:00 Agent OS Overview01:13 Model Access Concerns03:16 Posting Strategy Growth04:41 Local Models for Clients05:36 Managing Client Agents07:10 Automating Video Creation07:57 Fact Checking Workflow08:36 Windows Install Support08:54 Hermes Jarvis Voice Agent10:09 Paperclip Social Team11:57 Systems Over Models12:37 Hermes Desktop vs OS13:28 Using GLM in Claude Code15:21 SEO Blog Automation16:15 AI Game Development Demos18:09 Kanban Board Upgrade19:58 Join the Boardroom21:02 Easy Setup Testimonials22:14 Final Wrap Up

Hermes Agent Mixture of Agents (MoA) Presets: Build a Model Panel to Beat Gated Frontier ModelsHermes Agent has released Mixture of Agents (MoA) Presets, allowing multiple AI models to run in parallel and be aggregated into a single stronger answer, similar to systems like Fusion and Sakana Fugu. The script explains this as a workaround for gated or limited-preview frontier models (e.g., GPT 5.6 and Claude Fable 5) by improving performance through a “panel of experts” approach rather than relying on one model. It introduces Hermes Bench, noting claimed reference MoA scores 8% higher than Opus and 11% higher than GPT, and highlights top performance from an Opus 4.8 + GPT-5.5 aggregator. Viewers are shown how to update Hermes, select MoA presets via terminal and dashboard, configure via desktop app or config.yaml, and switch presets with provider commands.00:00 Hermes MoA Presets00:43 Why Models Are Gated02:09 Update and Enable MoA04:06 Commands and Config04:49 Top Hermes Bench Combos05:04 Panel Beats Genius05:47 Fusion and Alternatives06:35 Build Systems Not Models07:34 Agent OS Workflow08:42 Old Way vs New Way09:39 AI Profit Boardroom10:58 Wrap Up and Links

Fable 5 & Mythos-5 Update: Partial Lift, Limited Access, and Why You Need an Agent OSThe script discusses new updates on Anthropic’s Fable 5 and Mythos-5, citing reports that the Trump administration struck a deal allowing Mythos-5 to be released to a vetted group of 100 companies and some federal agencies, framing this as a partial lift after both models were pulled worldwide. It compares this gating trend to OpenAI’s GPT-5.6 limited preview and raises questions about whether access will be US-only and whether public access to frontier models is ending, especially as China releases open-weights models like GLM 5.2. The creator argues viewers should “focus on the system, not the model,” promoting an agent operating system that can swap between models and maintain workflows, memory, and prompts, and advertises the AI Profit Boardroom community and training.00:00 Mythos-5 Deal Update00:41 Gated Access Explained01:12 Negotiations and Timeline02:38 Agents Beat Model Gates05:14 Stop Chasing Models07:08 Agent OS Demo Highlights08:45 Boardroom Offer Breakdown09:38 Community Tour and Wrap