💻 **Tom Tunguz: Birds Don't Fly Like Planes — Why Local AI Models Win on Time-to-Answer, Not Tokens/Sec**

Benchmarking DeepSeek V4 against two local Ollama models on 25 VC tasks, Tunguz finds token speed is the wrong metric — **time to answer** is what matters.
• A local Qwen3.6-35B model generates **2.2x faster** than its Qwen3.8-27B peer, yet finishes slower because it "thinks" **3.1x** longer — quality across all three models ties at 8/9.
• Smaller local models reason more to close the knowledge gap with bigger cloud models, like a bumblebee taking a longer flight path to the same destination.
• _Takeaway:_ your laptop can now run a model as capable as most cloud options — just budget for latency, not throughput.

[🔗 Read more](/leaving?url=aHR0cHM6Ly93d3cudG9tdHVuZ3V6LmNvbS9iaXJkcy1kb250LWZseS1saWtlLXBsYW5lcy1uZWl0aGVyLWRvZXMtYWkv)
[#tomtunguz](/search?q=%23tomtunguz)