There’s a new open weights model out this week many are talking about - Kimi K3.
It’s not technically open weights YET, but on their launch blog, they said" “The full model weights will be released by July 27, 2026.”
At first pass, what stood out initially is the size of the model. K3 is a 2.8 trillion parameter MoE (16 of 896 experts active per token), which makes it the largest open model ever released (I think this is true, but haven’t fully fact checked everywhere). It’s ahead of DeepSeek V4-Pro at 1.6T (which I believe was the largest before K3). It has a 1M token context window, native multimodality, and seems to be built for agentic work. A lot of the launch materials / demos focused more on things like long-horizon coding, demos where it navigated massive repos, multi-hour autonomous runs, etc.
On benchmarks (Moonshot's own numbers, so take with grain of salt, the technical report isn't out yet, and you never really know what the “details” behind how benchmarks were run are), K3 beats Opus 4.8 on the majority of published evals, (but still falls behind Fable 5 and GPT 5.6 Sol). My TLDR - Kimi K3 is arguably the best open model released, sitting roughly one generation behind the frontier.
The other thing that stood out to me is the pricing. Historically, open weight models have come out and been a fraction of the pricing of frontier models. Looking at GLM 5.2, that model was $1.40 per 1m input tokens and $4.40 per 1m output tokens. Using a 80% / 20% input / output ratio (which many benchmraks standardize to), the “blended” cost of GLM 5.2 is $2 per 1m tokens. This compared to Opus 4.8 at $9 and GPT 5.5 of $10.
Kimi K3 pricing is $3 per 1m input tokens, and $15 per 1m output tokens. Blended price of $5.40. This is starting to get MUCH closer to “frontier” pricing vs “open weights” pricing. I think there’s an interesting question here - is this the trend of future open weights models? Will their pricing start to converge to frontier pricing? What happens if the price gap between open weights and frontier models starts?
There’s also another part of the pricing discussion I’ve discussed earlier. How token efficient are the models? GLM 5.2 was much cheaper, but it was way less token efficient. So for the same prompt it would use more tokens than a frontier model (by 2-3x). So the “headline” price understated the actual price to serve. I haven’t seen an analysis yet on the token efficiency of Kimi K3, but there’s a good chance when you factor in token efficiency the pricing is even closer to an Opus 4.8 or GPT 5.5.
BUT - there’s another (maybe even bigger) pricing consideration with Kimi K3. Everything I just described is API pricing (which for a closed model is all that matters). But K3 is (about to be) open weights and historically, that meant the API price was just one option. If you don’t like the pricing from Kimi, you can simply download the weights and run the model yourself (or go to someone like Baseten).
With a closed model, you just have to pay whatever the lab charges, and the price includes their margin. You don’t have an alternative. With an open model, the free license itself was partially what made it “cheap”: you could download the weights and skip paying anyone's margin at all, or the inference providers could all serve the same model and compete on price / performance.
Maybe said another way, open weights were a cost lever and the “license” was the discount. At 2.8T parameters, that math on that “lever” doesn’t exist to the same extent (or at all). The model was trained at MXFP4, so the weights alone are ~1.4TB. Meaning 10+ H200s just to load it, before you touch the KV cache on a 1M context window (napkin math so don’t hold me to it).
Moonshot is pretty up front about this as well. Their own launch blog recommends “supernode configurations with 64 or more accelerators.” The company releasing the weights is also saying that you need a 64+ GPU supernode to run them well. Which get’s into the next “hidden cost” - it’s as much a capex question as a self hosting one (and if you’re not self hosting, someone else like Baseten will, and they will have to run it on a larger compute footprint).
One force that historically factored into open model pricing (anyone can serve it) is much weaker when “anyone” means “anyone with a supernode and a serving stack tuned for a brand new attention architecture.” What we don’t know is what happens when they open the weights on July 27 - will 3rd party model servers undercut the $5.40 blended pricing? Will they be able to?
We don’t really know the true margins on Kimi K3, so it’s harder to comment on the true pricing until we’ve had a bit more price discovery from those serving the models. A 2.8T model probably has a structurally higher serving floor than a 1T model no matter is serving it. The K2-era 10x discount existed (partially) because those models were both smaller AND served at thin margins.
K3 only gives you the second one. For two years “open” and “cheap” were used interchangeably, but they were never the same thing. They were certainly correlated, because open models happened to be small enough that anyone could serve them. K3 is the first open model big enough to break that correlation. Which brings me back to the initial point - what is the true cost of open weights models, and how are they trending.
All of this to say - it’s very easy to make the big claim that “open weight models are going to kill the frontier! They’re way cheaper but just as performant!” However, there’s usually a lot more than meets the eye on these surface level arguments (however trendy they are to say on Twitter!)
Comments