@landontgreen It worked great but it ate so much extra memory over nvfp4 and exl3 that my context window was effectively useless. If I can get that sorted out I'll happily stick with tensor fold it screamed in decode
@mfranz_on The 4b model I run on my phone is such garbage compared to 27b. But the gap between 27b and something like 300b models isn't nearly as large
@ItsmeAjayKV If the price is 2x but the performance is 4x its a net gain. Id be a bit annoyed at the price but if I can cram a full sized GLM in it ill take it
@BenNapier Yea this shit is nothing more than a future ban on open source and letting people have their own models. Don't worry the big companies will "generously share" their old models and keep the best for the elites
@MikeBradleyAI Until Deepseek v5 250b comes out and everyone is saying memory is all that matters. Its a cycle and with 3.8 27b we are back to the TPS arguements
@FrankBr05713205 The coolant pipe on my Mercedes was plastic and cracked. The replacement was stainless steel. Why didnt they just do that shit to begin with
@roboindexio I mean if the driver didnt have to exit the vehicle it would be better. But he has to get out and open the door and load the boxes anyway. He could just take the boxes 20 steps just as quick
@picocreator Ive thought of robotic arm on a rail for the garage/workshop for a long time now. Being up high keeps it out of the way when it needs to be