Why does it actually matter that DeepSeek and Kimi are both hitting open weight milestones this same week?

[Lewis_43]

Quick explainer for anyone confused by why this specific week keeps coming up in AI discussions. DeepSeek V4 Pro actually shipped its open weights back on April 24, available under a fully permissive MIT license on Hugging Face from day one, at roughly 0.44 dollars per million output tokens, easily the cheapest capable frontier grade open model on the market. What is actually happening this week is that DeepSeek's older legacy API endpoints, the original deepseek-chat and deepseek-reasoner names, are being retired on July 24, forcing anyone still using those older endpoints to migrate over to the V4 naming

Kimi K3 is the genuinely new part of this story. Moonshot released K3 as an API only model on July 16, and it reportedly topped a major coding leaderboard almost immediately, but the actual downloadable weights were not included at launch. Moonshot has committed to releasing those full weights by July 27, under a modified MIT license that only adds an attribution requirement once a company crosses 100 million monthly active users, which is irrelevant for the vast majority of businesses that might want to use it

Why any of this matters practically, once K3's weights are actually out, any company with enough hardware can self host what is reportedly a near frontier coding model with zero per token API cost, rather than paying Moonshot's API pricing which currently runs 5 to 17 times higher per token than DeepSeek's V4 Pro. That is the real headline, not the specific date, but the fact that self hosting a competitive coding model becomes an option for teams who previously assumed only the expensive closed API providers could deliver this level of quality

Kimi K3 Open Weights: A July 27 Readiness Checklistwww.digitalapplied.com

[Seb51]

So DeepSeek V4 has actually been open for months already, it is just the old endpoint names getting deprecated this week, that clears up a lot of confusion I had

[Jade]

The self hosting angle is really the whole story, paying per token forever versus a one time hardware cost changes the calculus completely at scale

[FairDos47]

K3's pricing being 5 to 17 times DeepSeek's while weights are not even out yet is a wild gap for what is supposedly a similar tier of model

[Sandworm]

Curious whether Moonshot actually hits the July 27 date or if it slips, plenty of these announced dates end up being more like loose targets

[GlassyCandle]

That attribution clause only kicking in above 100 million monthly users is basically a non issue for 99 percent of companies that would use this

[Sinead77]

Good breakdown, I did not realize deepseek-chat and deepseek-reasoner were being sunset, definitely need to check what my own scripts are still pointing at

[Optimiser Ruby]

Self hosting a 2.8 trillion parameter model is not exactly trivial either, you need serious hardware to actually run this yourself

[Matthew80]

This is a good reminder that open weights and open API access are two very different things people often conflate

[RightAbout]

The real test once K3 weights actually land will be independent benchmarks, company published numbers always look better than reality

[Josh_79]

Appreciate the plain explanation, the pricing comparison alone makes it obvious why enterprises are watching this week closely