Open Source Models Approach Closed Source Performance

You need to start experimenting with open-source models right now if you haven't already. Until just a few months ago I would have said that it's optional, but the latest releases have been eye-opening. Learn how to set them up, learn the benefits and drawbacks, compare the cost for the throughput that you need, come to your own conclusions. It's hard to trust benchmarks, but Minimax M2.5 does feel close to SOTA, which should be terrifying to anyone working at an AI lab in the United States (unless they have open-source models in the works that the rest of us don't know about). And you can run at 100tps for an hour for under a dollar. Personally, I don't even care about benchmarks. I care about the graph below. There's a chance (slim chance, but still) that the next open-source model is BETTER than the best closed-source in the US. Congratulations to the MiniMax team... someone cooked here.