Yuchen Jin at Databricks flags GLM-5.2 as the point where open-source models deliver Claude-level performance for coding and agent workflows, driving companies to download weights they can fine-tune and keep. The 744B-parameter MoE model from Z.ai ships under MIT with a 1M-token window and tops several open rankings while trailing Claude Opus 4.8 by narrow margins on long-horizon benchmarks.
GLM-5.2 is the open-source Claude moment.
The demand we’re seeing at Databricks is astonishing. The world is going to see massive adoption of oss LLMs.
Also, more companies will shift toward post-training their own models on top of oss models and owning the weights.
Databricks reports extraordinary uptake of open models that let enterprises avoid API rate limits and own their post-trained checkpoints. Inference optimizations now push the model past 390 tokens per second on the platform.
The rapid enterprise shift toward fine-tunable open weights, Marcus argues, shows closed frontier labs no longer hold durable advantages once performance gaps narrow this far.
Many users praise GLM-5.2 as a strong frontier open-source model delivering agentic coding performance and capitalizing on US restrictions for wider adoption, while a few criticize specific replies or question the value of post-training it.
No Digg Deeper questions have been answered for this story yet.
GLM 5.2 is the ultimate culmination of what i have argued here for the last three years. no technical moat and widely known formula -) convergence -) price wars -) small or negative margins.
GLM-5.2 is the open-source Claude moment.
The demand we’re seeing at Databricks is astonishing. The world is going to see massive adoption of oss LLMs.
Also, more companies will shift toward post-training their own models on top of oss models and owning the weights.
@Onwuta_Kelvin as long as the worst case doesn’t happen
@aaryan_kakad Performance always vary depending on the workload. We keep working on optimizing it for our customers’ workloads, which can differ case by case.
@Yuchenj_UW Sometimes with every disappointment comes an opportunity. US restrictions have indirectly propelled oss use and adoption. We are witnessing another big change in AI landscape
@Yuchenj_UW the GLM 5.2 in databricks is what quantization and what speed currently?
@Yuchenj_UW This reads like a lot of validation.
@Yuchenj_UW @Onwuta_Kelvin Or the gov bans huggingface and inference providers from hosting OSS models and makes it harder to access them. This is more worse cause not everyone can run local models that are mythos tier.
@mithilyaganti @Yuchenj_UW Trust me, it gets better if they do this
In China there's a joke where they refer to Trump as "builder of the Chinese nation" (because his policies end up unintentionally hurting the USA and helping China) and this just seems like another example. By meddling with the AI model releases in such a ham-handed way the government is making Chinese open source models look a lot more appealing.
@Yuchenj_UW To think this is just a 744B parameter model, imagine what will happen when 1.6T parameter Deepseek-4-pro launches in future Ds-4.1 & Ds-4.2 versions with improved performance and maintain it current cheap inference pricing 😄
@Yuchenj_UW That's real , as you can see on Morpheus Infrence Marketplace http://App.mor.org
@Yuchenj_UW Exactly 💯. I just called this a few days ago too. Must be the natural conclusion for businesses after taking it in the a** from Claude and ChatGPT this week🤣
@Yuchenj_UW thinking you're right
@Yuchenj_UW Not moving until RedHat comes with RH Enterprise GLM 5.2 :D
@Yuchenj_UW I hope next version of GLM can do more optimization on the agentic task
@Yuchenj_UW Curious what demand are you referring to? Are you reselling LLMs at Databricks?
@Yuchenj_UW @stevehou @prlnet
@Yuchenj_UW Why does anyone use a proprietary closed produce like databricks?
kimse bahsetmiyor ama GLM-5.2'nin açık kaynaklı Claude anı olması, aslında büyük teknoloji şirketleri için korkunç bir kabusun başlangıcı. yıllarca süper bilgisayarlarla interneti ezberletmeye çalıştılar, milyarlar harcadılar. şimdi 15 milyon parametreli, tek gpu'da çalışan ufacık bir model fizik kurallarını öğreniyor.
bu resmen bir "dejavu" anı: tıpkı 2000'lerin başında linux'un microsoft'u terletmesi gibi. sadece bu sefer, "linux" makine öğreniminin ta kendisi. yakında "bizim de bir açık kaynak stratejimiz var" diyen ceo'ları görürsek şaşırmayalım. zaten birçoğu multi-llm stratejisine geçmişti bile.
@Yuchenj_UW the llama moment but for the frontier
every enterprise that's been waiting for open source to catch up just got their answer