the difference

I'm gonna start out this post with an important point: openweight models are not used the same way as proprietary models. At the time of writing this post the current SOTA (state of the art) models by OpenAI and Anthropic are GPT 5.5 and Claude Fable 5 respectively. The latter only having been released a few hours before writing this post and not fully being benched and tested yet. For these the use case is basically just to always use the latest and best model and old ones are quickly deprecated. Openweight models are used differently; you don't just use the best model always if you did that you might as well just use the OpenAI models.

how to actually use openweight models

Unfortunately models come out every few weeks these days so this post will get outdated quite quickly. Nevertheless right now the up to date openweight models are Kimi K2.6, Minimax M3, Deepseek V4 and GLM 5.1. Of these in my experience the very best to use for most cases when you just need working code is Kimi K2.6 and that is what I usually go with when I just need some fast working code. This is not the only model to use though since it is also quite expensive compared to some others, my personal favourite for model usage to not use for coding has to be Deepseek V4 flash. It is a super cheap model and most use cases are only a few cents and compared to other super cheap models actually is quite usable. If you were to try using local LLMs or other small models like Nemotron 3 ultra it is much better. The thing is it does call tools properly and reason nicely, while these other cheap models will sometimes just stop in the middle of a task or fail tool calls over and over. I've also used Minimax M3 before for some reverse engineering and cybersecurity tasks and it does a good job for a nice price, it's not quite at the price of Deepseek V4 flash but it also isn't as expensive as K2.6. I personally wouldn't feel as comfortable using it for prod code as that model though.

why to use openweight models over the giants

The actual reason to use openweight models is not that simple. The obvious things are speed and price since the big labs are compute limited and those are the first things to go. The other thing is reliability: proprietary models by the big labs are hosted by just them, openweight models you could run yourself, that means that the alternate hosts will still have them in case they were to take them down, this means that if you use an LLM proxy such as OpenRouter you will most likely be good on great uptime and them staying there. Another thing is having an alternative, you can use whatever host you want or run the models yourself and this means that if you don't want to support certain companies or countries YOU CAN.

the downside with openweight models

Openweight models are awesome I think I made it clear that I find that the case but there are still downsides to them that cannot go unmentioned. The first thing is that they are just worse, if you look at benchmarks such as SWE-bench you would see that some openweight models score scarily close to some of the proprietary models from the big labs. The issue is that this benchmark is unfortunately not really that usable anymore, it is in training data sets for models now and some of the Chinese labs with the openweight models actually do post-training on it to get better benchmark scores. If you look at a benchmark such as DeepSWE instead you see a new benchmark with values which I believe better reflect the actual usability for coding for these models and as you see it is not even close. I think some models are a little too low such as K2.6 but models such as Deepseek V4 pro at 8% I fully believe.

my personal usage of openweight models

I've personally played around and used openweight models a good bit. I have a subscription for Opencode Go to get most of the models for cheaper prices and in case they are not available there I will also happily pay API prices on OpenRouter or directly at the provider to use them. I have used Kimi K2.6 to write considerable code for some of my projects and as I wrote before also Minimax M3 for some reverse engineering. The Deepseek V4 family stays one of my favorites for cheap fast testing and I've not used them for writing real good code but I recently used Deepseek V4 flash in Hermes agent for some research and it went really well since it does call tools really reliably.

the future of openweight models

As of writing this post GLM 5.2 is to be released very soon as the API already shows that it exists be that by accident or not. How it goes after that I don't know. I hope that these models do get as good as the proprietary models such as GPT 5.5 of course for cheaper and faster. Even though this would be awesome it's also really scary since the models don't have the option of being filtered for cybersecurity later on. On API the lab can obviously do that but if the weights are released it is too late. If there is a model for cheap that can hack anything without any censoring I don't want to be the one pushing hotfixes.

conclusion

In conclusion I hope we can all agree that openweight models definitely have their uses if you don't want to pay $25 per million output tokens or want more than 50 TPS but if you want the bleeding edge, the SOTA the models that write the best code you will have to cough up this money or pay for a subscription.