Flyte enables production-grade orchestration for machine learning workflows and data processing created to accelerate local workflows to production.

Flyte

<https://huggingface.co/blog/zai-org/glm-52-blog>

This is by far the best open weight model that i have seen till date

image.png

Jun-18-2026 16-07-58.gif

I just served a quantized version of this on our demo Union cluster!
<https://unsloth.ai/docs/models/glm-5.2>

This is in a `g6e.12xlarge` instance, which has 192Gi of VRAM… which means only 60_000 token context window because it can’t handle any more than that :sweat_smile:

GLM 5.2 index sharer is something new that helps with “needle in a haystack” problem but i think the size is large, api calls does not working for many people and use cases.

Next big thing would be a smaller model with higher accuracy. I should be able to run it on my watch, if required, without internet

<@U01DYLVUNJE> Which quantization scheme are you using? Are you doing a k/v cache quant as well ? Is it giving good result for long running agentic task?

You can run it on a spark device (dual pack = 256 GB VRAM) GB10, quantized 2 bit, 1M context window, No k/v quant.