Open-weight or API? How to decide when the break-even numbers disagree
This summer brought a run of large open-weight models. Moonshot AI released Kimi K3, a 2.8-trillion-parameter model, on 16 July 2026. Z.ai’s GLM-5.2 (753 billion parameters, reported under an MIT license) and Alibaba’s Qwen 3.8-Max (2.4 trillion) arrived in the same stretch. One report says open-weight models have narrowed the gap with proprietary systems to single digits on most benchmarks that matter to enterprise buyers. So the question gets asked a lot: should we host one ourselves?
The cost question, and why the answers disagree
API prices vary widely. One comparison lists DeepSeek V4-Pro at $0.87 per million output tokens and Kimi K3 at $15, a gap of more than 17 times. That alone changes the sums. But published break-even points for self-hosting don’t agree. A developer write-up on Kimi K3 puts the break-even at roughly 2.7 billion output tokens a month. CoreWeave, which sells GPU infrastructure, says cost analyses this year put it near 5 to 10 million tokens a day for a right-sized open model.
Figure 1 · Two published break-even estimates
A spread this wide says the answer depends on the model you pick, the hardware you can keep busy, and what you count as cost. Do your own sums with your own volumes.
A way to decide
Figure 2 · Choosing a path
| Question | Leans API | Leans self-host |
|---|---|---|
| Volume | Low or uneven | High and steady, so the GPUs stay busy |
| Data rules | Provider terms are acceptable | Data must not leave your network |
| Team | No one to run GPUs | You can operate and monitor the stack |
| Model needs | Best quality on hard tasks | A smaller model is good enough |
| License | Not your concern | Read it. Reports on K3’s terms conflict |
One practical point: a self-hosted model only saves money when you keep expensive hardware busy around the clock. And licenses differ. Some sources call Kimi K3’s license MIT, others describe a custom license with conditions for hosting providers, so read the actual text before you plan around it.
It doesn’t have to be either
A mixed setup is often the sensible one: fast, cheap models for intake and classification, and a larger model, hosted by you or by a provider, for the hard reasoning. That is a routing problem, and we covered the thinking in routing belongs in the harness.