Recursive Flint: This OPEN Local Coding Model is CRAZY! (Tested with Pi, OpenCode & PandaCode)

Join this channel to get access to perks: https://www.youtube.com/channel/@aicodeking/join In this video, I'll test Recursive Flint, a new open mixture-of-experts coding model with 40 billion parameters in total and only 2.5 billion active per token, built to run agentic coding on your own computer. I'll go over Recursive's claims, set it up on my M4 Max with llama.cpp after their own rec engine refused my Mac, and measure how fast it really is at short and long context. Then I'll show you how to use it in PandaCode with Pi and OpenCode, and put it through four real tasks: a landing page, a bug fix, a backend API and a Breakout game. I'm not testing it on KingBench, because it's a small local model and that wouldn't be fair, so I checked every result myself instead. -- Key Takeaways: 🚀 Recursive Flint is an Apache 2.0 mixture-of-experts model with 40B total parameters and about 2.5B active per token, made for local agentic coding. 📊 Recursive claims it beats every MoE model under 50B on SWE-Bench Pro, Terminal-Bench and SWE-Bench Verified, and runs over 125 tokens per second on an M4 Pro. ⚠️ Recursive's own rec engine only supports the M4 Pro, M5 Pro and M5 Max, so it refused my M4 Max and I used llama.cpp instead. ⚡ In llama-bench it did about 110 tokens per second on short prompts, but about 52 at 8K and 32 at 32K context. 🧠 Coding agents send big prompts (about 17K tokens from Pi and 26K from OpenCode), and in one long run the speed fell to about 9 tokens per second at 115K context. 💻 Flint works in PandaCode through Pi and OpenCode, and it shows up in the model picker once you add the config. 🔧 If you use OpenCode with a local model, turn off the provider header timeout, or long prompts make it give up after five minutes. ✅ The bug fix (7 of 7 tests) and the link-shortener API (14 of 14 tests) were both correct when I checked them by hand. 🎯 The landing page looked good apart from huge pricing checkmarks, and the neon Breakout game worked well. 👍 It's not a frontier model for big tasks, but for free, private and offline coding on focused tasks it's a really good option.