@ryanallis9CEO, AI Lounge & SaasRise
Recursive releases Flint, a 40B open-weight coding model that runs on laptops
Is this going to kill Anthropic's revenue from Claude code? Will we all just run local coding models on our own hardware in a year or two.
agentic engineering
Comments2
I run Ollama locally and it's great for small stuff, but for real multi-file work I still reach for Claude Code. I think local models eat the cheap, simple tasks first and the hard agentic work stays in the cloud for a while. Curious if Flint's SWE-bench numbers hold up on messy real repos.
I haven't benchmarked this one but it looks like it was distilled on Gemini. The cost of hardware has now inflated where subscriptions actually make sense. Unless an intelligent individual or company increases the intelligence density of these models to run on lower end or lower capacity (VRAM) hardware. Infact the new laptops for 128gb which is what you need to run frontier class models are going to be $3k+. So you'll run basic computer use, writing documents and emails locally. To do actual useful engineering or design work you will use frontier. Not to mention ease of use. Right now the steps to get a local model running are greater than buying a frontier ai subscription you can use in your browser. Businesses at the small and mid level are still using AI for writing documents instead of ACTUAL work but most people are too dumb to install local AI. This is the sweet spot for frontier labs (low usage with enterprise level subscriptions)
2 comments
I run Ollama locally and it's great for small stuff, but for real multi-file work I still reach for Claude Code. I think local models eat the cheap, simple tasks first and the hard agentic work stays in the cloud for a while. Curious if Flint's SWE-bench numbers hold up on messy real repos.
I haven't benchmarked this one but it looks like it was distilled on Gemini. The cost of hardware has now inflated where subscriptions actually make sense. Unless an intelligent individual or company increases the intelligence density of these models to run on lower end or lower capacity (VRAM) hardware. Infact the new laptops for 128gb which is what you need to run frontier class models are going to be $3k+. So you'll run basic computer use, writing documents and emails locally. To do actual useful engineering or design work you will use frontier. Not to mention ease of use. Right now the steps to get a local model running are greater than buying a frontier ai subscription you can use in your browser. Businesses at the small and mid level are still using AI for writing documents instead of ACTUAL work but most people are too dumb to install local AI. This is the sweet spot for frontier labs (low usage with enterprise level subscriptions)