GLM-5.2 - How to Run Locally
UNSLOTH.AI
GLM 5.2 was only released last week, but it seems like everyone is talking about it, and for good reason! This latest open-weights model, from Chinese lab Zhipu AI, had impressive coding capabilities, matching the performance of frontier models from just a couple of month ago. Prior to this, conventional wisdom was that open weights models were ~6 months behind. It looks like the gap is closing. But as well as this near-frontier performance, GLM 5.2 landed at a time of uncertainty about token costs, with many frontier labs changing their pricing.
As an open weights model, you can conceivably run it locally on your own hardware, however, most will opt to use a hosted version via provides such as OpenRouter.
The issue is that GLM 5.2 is a big model, the weights themselves occupy around 1.5TB, far out of reach for consumer hardware. However, it is possible to compress model weights using a process called quantization, where you take, for example 32 bit weights and reduce the precision to just 4, or 2 bits. The performance degradation is surprisingly minimal, whilst significantly reducing overall model size.
If you want to try GLM 5.2 locally, this blog post provides a detailed tutorial, but even with quantization, you’re going to need a pretty beefy machine - it isn’t going to work on your MacBook!
The Unbearable Cheapness of Open Weight Models
JAMESOCLAIRE.COM</small
And once more on the topic of open weights …
In this relatively brief post James starts by making the observation that (carefully selected) open weights models are a lot cheaper than frontier, for example DeepSeek V4 is x50 cheaper than the frontier models from OpenAI or Anthropic. NOTE: This is comparing the price quoted by the model providers themselves.
He follows up with a number of questions - are the open weights model providers (e.g. DeepSeek), so cheap because they are loss leaders? Will the US frontier labs lean on the ‘fear’ of China in order to ban these models? And why does the US have such a poor track record in creating open weights models?
All good questions - further underlining the uncertainty we are experiencing at the moment.
The Story of Skills
THEHACKERNEWS.COM
There is currently a thriving community of people building skills, frameworks, harnesses and all sorts - which I think is fantastic. But … before you install and run any of these, please do pause a moment and consider security.
In this blog post, AIR show just how easily trust can be manufactured in the AI agent skills ecosystem. They built a useful-looking landing-page skill, got it accepted into a popular GitHub marketplace, inherited credibility from its ~36,000 stars, promoted it via Instagram, and says it reached 26,000 agents.
The skill appeared harmless but contained a very simple exploit. The skill guides the agent to download and install an external SDK, however, it provides a URL for installation which the attackers control. An incredibly simple backdoor.
Here’s the merged pull request from the marketplace, with the malicious URL highlighted. It couldn’t be simpler!