It's still worth exploring GLM 5.2. Just a couple weeks ago, before Kimi K3 overshadowed it, GLM 5.2 was the reigning champion of open source models. The truth is, it's still an extremely strong contender - right around GPT 5.5 and Opus 4.8 quality, at about 1/5 the cost of Kimi K3. Clearly, it's still useful for many types of complex work, especially since it also performs faster, and has better hallucination metrics than Kimi K3.
I'm trusting Kimi K3 for really deep, complex, novel technical writing, but I'd expect GLM 5.2 can keep up in many software development tasks.
In fact, I now generally accomplish everything local with Deepseek v4 Flash, then have the monstrously big models perform final revisions, if/when there's any potential for concern that ds4f might have made an error anywhere.
I've been really surprised that the review of ds4f's work by those other models is generally extremely positive (most often adding cosmetic improvements and little stylistic fixes to the work of ds4f). So, ds4f does 95%+ of the important token generation work, and the other models consume 95%+ of the rate limited usage - and the thing I keep finding is that the huge frontier models just continue to validate exactly how solid ds4f is.
To be clear, for most practical purposes related to getting local work done, ds4f has become my fully trusted first choice, and for big development projects, ChatGPT with the zip file routine is still my go-to.
When ChatGPT and the other providers go out of business, or are forced to charge much more after investment capitol subsidies dry up, I'll expect to do most work with models in the class of ds4f, and use big models like GLM 5.2 to supervise and correct any critical work, if/where necessary.
GLM 5.2 is additionally interesting because https://github.com/JustVugg/colibri got it running successfully on normal desktop computers with only a CPU and modest RAM. Of course, that Colibri version runs at an extremely slow 0.05-0.1 tokens per second during cold generation (1 word every 10 to 20 seconds), on modest machines without a GPU, but it does get a bit faster with use, and it does work.
Colibri could only possibly be practical for overnight or longer generations, but it's truly incredible that a world class model can actually run on what would otherwise be considered useless hardware for LLM generation - hardware that would otherwise require tens-hundreds of thousands of dollars. I hope the Colibri project points to similar possibilities with other smaller MOE models, which may be able to run at a usably fast speed on machines with only a CPU.