600tok/s single request on qwen3.6 35ba3b with Ninfer on an RTX Pro 6000. Anybody remember that Comcast ad "stupid fast"?

A user achieved 600 tokens/second on a single request using the qwen3.6 35ba3b model with Ninfer on an RTX Pro 6000. While acknowledging it's not the most intelligent model, they find it suitable for "read+find" or coding tasks where brute-force processing is acceptable. Despite potentially using 20 times more tokens, its speed surpasses many other local models, making it a "stupid fast" and enjoyable tool.

This report highlights a specific model, qwen3.6 35ba3b, achieving 600 tokens/second, a speed described as "stupid fast" compared to many other local models.