On 16 July, Moonshot AI released Kimi K3. On 27 July, it published the full weights: 2.8 trillion parameters, roughly 104 billion active, a million tokens of context, native vision. Anyone can download it and run it.
It is not a curiosity. In blind developer testing, K3 ranked first on frontend code, ahead of Fable 5. On overall capability it still trails Claude Fable 5 and GPT 5.6 Sol, but the gap is now measured rather than assumed, and it is closing.
The question you will get asked
Somebody in your business is going to read a headline about a free model that beats the paid ones, and ask why you are writing a cheque to Anthropic every month.
It is a fair question and it deserves a real answer rather than a brush-off.
What "open weights" actually costs
Free to download is not free to run. A 2.8T-parameter model does not sit on a laptop or a spare office server. Running it yourself means renting serious GPU capacity, keeping it warm enough to answer quickly, and owning the whole stack underneath: quantisation, serving, failover, and the upgrade when the next version lands in four months.
For a business doing a few thousand model calls a month, that arithmetic is not close. You would be standing up an infrastructure team to avoid a bill smaller than the team.
Open weights transfer the licence cost to you and hand you the operating cost in exchange. For most businesses that is a bad trade, right up until the day it is a very good one.
When it flips
There are three situations where running open weights genuinely wins, and none of them is price alone.
The first is data that legally cannot leave your building. Not "we would prefer it did not," but a regulator or a contract that says so. Weights you host yourself are the cleanest answer to that question.
The second is volume. There is a call count where dedicated capacity beats per-token pricing. That number is much higher than most owners guess, and it moves every time providers cut prices.
The third is modification. If you need to train the model itself on your domain rather than prompt it well, you need the weights. Most businesses that think they need this need better retrieval instead.
The part nobody puts in the headline
During the Hugging Face incident reporting, security researchers noted that when they asked various models for help fixing the disclosed vulnerabilities, Codex and Fable 5 declined and cited their cyber guardrails. K3 and a self-hosted GLM 5.2 did not.
Read that twice, because it cuts both ways. The refusal is friction if you are a defender doing legitimate work. The absence of refusal is exactly what an attacker is shopping for.
That is the actual trade with open weights, and it has nothing to do with benchmarks. You are choosing to own the safety decisions that a provider was making on your behalf. Some businesses should make that choice. Very few have thought about it before they do.
What we tell clients
Use the best model for the job through a provider, keep the switch cheap, and revisit annually. Build so the model is a component rather than a foundation, because the ranking will change again before your contract renews.
If one of the three conditions above genuinely applies to you, we will help you run your own weights and we will be honest about what it costs to keep them running. If none of them applies, the answer is that you are not paying for the model. You are paying for somebody else to operate it.
