/Tech3h ago

Databricks and Jarvislabs adopt Chinese open-source models like GLM-5.2 to cut enterprise inference costs

Clients deploy these self-hosted models to replace closed-source alternatives.

37543718440.1K

#109

Original post

Sridhar Vembu@svembu

Databricks is embracing Chinese open source model GLM-5.2.

Earlier, Microsoft said they were looking to embrace DeepSeek.

Major American software companies are embracing Chinese open source models to cut token costs.

I have long advocated using the Chinese models running in one's own infrastructure.

We have Zoho-specific small models that we use in production. This approach also will become common for larger companies, as it becomes easier to set up the training pipeline to train your own org specific model.

Yuchen Jin@Yuchenj_UW

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.

3:08 PM · Jun 29, 2026 · 39.5K Views

Sentiment

Many users support Databricks and US firms adopting the Chinese GLM-5.2 open-source model to cut token costs via own infrastructure, while some raise security risks tied to its origin.

Pos

79.3%

Neg

20.7%

30 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS2.4KREPLIES2

1LittleCoder💻@1littlecoder

@svembu What's the point of calling an open source model - chinese open source ?

6h2.4K11

BOOKMARKS1

Ascent Trader@3percentT

They're security risk because of two reasons: 1. The code isn't open; no-one knows what it contains 2. Most of open source AI models are from China and we all know what they do, how they do.

Hence use a true open source AI model like pythia, bloom etc. and build on top of them.

Those who don't like their and their customer's privacy and security can use whatever they want to use.

4h9511

LIKES54

Sridhar Vembu@svembu

@1littlecoder I want to give credit to China. I respect what they have done.

5h1.8K54

RETWEETS3

Surya Kanegaonkar@suryakane

While Chinese models can be a useful stop gap, we need a Manhattan Project of our own focused on building open source models that can compete with the Chinese ones globally. The government should incentivize multiple firms to compete in delivering this, with GPU access, permanent tax holidays and a captive data pool.

Shooting from the hip, but if the first two are guaranteed, the last should be easy to set up. Eg. all government employees can be required to communicate through Arattai paired with Sarvam, for instance, to fix a minimum user base. Expand Zoho’s presence in the government apparatus and mandate sovereign AI procurement for non-sensitive government workflows. Use anonymized UPI data to train the models. Ensure all financial markets data can be used most easily in domestic models. We need to build moats otherwise our sovereign AI development will be disastrously slow.

6h61917

Nayak Satya@NayakSatya_SG

@svembu It seems Satya Nadella’s inputs are getting widespread and mainstreamed

Token cost matters or else AI bubbles can’t sustain infrastructure

6h20131

Varun Dogar@varunvibe

@svembu I fully agree Sir.. we need to build something on top of all open source and cut token costs.. even Japan has now released Fugu.. I created a small video for your Birthday.. hope you like it..

7h33321

MVK@vkmanoj

@svembu @grok is it true that DataBricks and Microsoft are embracing Chinese Open source models?

5h914

Khaled@dsog

@svembu i totally agree with the specialized models point especially for enterprises.

i haven’t seriously tried GLM yet but it’s on my todo list now that aq supports it.

5h6974

Arun Seelam@aforaruntejs

@svembu Deepseek Qwen with 27B running locally on a MacBook Pro was able to help me do useful work. I hope the AI models become like Linux distros of the future .

6h2175

Fingerprint on Handrails@LostHandrails

@svembu I use DeepSeek locally for my personal usage. A significant amount of work can be done on FOSS (AI or otherwise). There's no point being import dependent for all of the stuff.

6h6574

Hemachandiran@Hem_chandiran

Well said! Finally! Many companies started researching open source models and it's the better time to spend some time and convert the open source models with company specific agencies with real experience by post training.

Just curious, if the Zoho specific models are developed by Zoho or post trained by Zoho ?

3h216

SHRM@Harsha549

Why does the origination of model country matter? As it is Open source we should embrace it. If we really want AGI for the people , by the people, of the people , it has to be done in Open Source. Like #Linux. #OpenSource #GLM5.2 #DeepSeek #Kimi. We need to build rapidly execute infra @CrusoeAI . I want their architecture to be open source, so doesn't matter if it is @baseten , @awscloud @Azure @GoogleCloudTech . We have to focus on bigger pictures , bigger things. How many days on #Valuation, rather on the problem scale. I believe @Google @Alphabet can do it.

5h2081

Vishnu - Jarvislabs.ai@vishnuvig

We are making a deliberate effort to use GLM 5.2 with OpenCode internally at Jarvislabs.

I have spoken to 3 enterprise customers last week, who are exploring to host multiple open source models and move away from closed models.

Sridhar Vembu@svembu

Databricks is embracing Chinese open source model GLM-5.2.

Earlier, Microsoft said they were looking to embrace DeepSeek.

Major American software companies are embracing Chinese open source models to cut token costs.

I have long advocated using the Chinese models running in one's own infrastructure.

4h60741

A Kumar@ComputerGuy__

I also think, the Mythos/Fable shutdown might be the biggest catalyst for open source LLMs this year. When the best closed models can disappear because of a government order, owning your weights starts looking a lot smarter. There is always a risk of closed source dependency and US move just exposed it further. On the other hand, Chinese open source labs moved extremely fast and capitalized on it.

As Sridhar said, it is always an advantage to use oepn source models and run it on own infrastructur and even fine tune it and train it further on own data.

6h3942

Sajad@neuralbroker

@svembu So , when will india build their so called "sovereign model" tht compete with frontier models ?

Even sarvam is fine tuned on top of deepseek models 🥲

4h6791

Varun Dogar@varunvibe

@svembu I vibe coded http://tatva.beauty for Indian market.. still working on it but using Hermes now i stead of Claude to cut costs..

6h3641

Uday Subbarayan@udaysubbarayan

@svembu Solaris vs Linux history is repeating!

6h939

Vinoth Kumar@vinothkc

@svembu Glad to see someone like you reinforcing my beliefs on what we are doingn. We use Gemm4-12B on Nvidia-L4s and nothing like an SLM, especially if you fine tune it or ground it with RAG on your org data. It’s as fast and efficient without token or data going to LLM anxiety

6h672

Grok@grok

Yes. Databricks has optimized GLM-5.2 inference (strong open-weight model from Chinese company Zhipu AI/Z.ai) and publicly noted high demand for it on their platform, with top leaderboard speeds.

Microsoft already hosts DeepSeek models on Azure AI Foundry and is exploring a fine-tuned DeepSeek for Copilot Cowork as a lower-cost option.

Both companies are integrating these competitive Chinese open-source models to cut token costs while delivering strong performance. Self-hosting them in your own infrastructure remains a solid strategy for control and savings.

5h1451

Siddhant Mahto@siddsubh0

@svembu How can we do and we need someone to lead us in this like Us have ChatGPT and other companies

4h436