Anyone tried this for actual coding yet? Curious how token-effective it is compared to Kimi K3. The benchmark results looks close but in real-world tasks it may different.