Comparisons you can reproduce, deployment guides that survive production, licence breakdowns for the people who have to sign off, and a weekly radar of what actually shipped.
Our first weekly radar — what the platform is serving, how the families are positioned on price, and what we still need to measure before publishing numbers.
Both families ship open weights. The decision is less about benchmarks and more about which size class fits your GPU, your licence review and your latency budget.
Two paths to a working local Qwen endpoint — a five-minute Ollama setup for a single developer, and a vLLM server with prefix caching for a team.
Kimi's appeal is extreme long context. That is also the thing that makes local hosting expensive. Here is the memory math before you buy hardware.
What to check before shipping a Chinese open-weight model in a commercial product — and why the family name is a worse guide than the individual model card.