Meta has just released Muse Spark Open, an open-source giant with 816 billion parameters, yet its benchmark performance has been surpassed by Qwen3.8-Flash-Next, which has only 18 billion parameters. According to the latest real-world testing by third-party firm Artificial Analysis, Meta’s massive model scored just 39 points overall, while Qwen3.8 achieved 40 points. Even more striking is the difference in activated parameters: Meta’s model requires activating 42 billion parameters per inference, whereas Qwen needs only 6 billion—seven times fewer. The cost of local deployment is even more apparent. After quantization to 4-bit, Meta’s model file weighs a staggering 506 GB, making it nearly impossible for ordinary servers to handle; Qwen, by contrast, is only 111 GB and can fit entirely on a single professional GPU. Despite using over four times the parameters and VRAM, Meta’s model has been outperformed in inference capability—big tech’s current approach of blindly scaling up parameters now demands serious reevaluation.