They are the future as long as someone is releasing them. From what I read lately it could be that OpenAI and Anthropic will get the regulation they caved for and unexpectedly China will stop releasing good open models.
they're presentish, if you prepped pre Fall 2025. The current prices for local model GPUs is insane.
I picked up the AMD Strix Halo with 128GB a year ago, and had my company pick up a framework one prior to that.
Qwen3.8-Flash-Next is the equalizer. I'm building autonmously with it for personal and work. There's a halogen-flash-server that runs rings around llamacpp.
With the right engineering harness, I'm keeping the same agent busy through 6 million tokens by recycling the context with dynamic context pruning and eviction so it's able to just keep chugging with a little direction & decisions.
10 comments
[ 0.23 ms ] story [ 10.6 ms ] threadOpen weight models are the future.
I picked up the AMD Strix Halo with 128GB a year ago, and had my company pick up a framework one prior to that.
Qwen3.8-Flash-Next is the equalizer. I'm building autonmously with it for personal and work. There's a halogen-flash-server that runs rings around llamacpp.
With the right engineering harness, I'm keeping the same agent busy through 6 million tokens by recycling the context with dynamic context pruning and eviction so it's able to just keep chugging with a little direction & decisions.