Current AI stack

@tys.ais

3.0K followers · usually 1.1K views a post

Views

38.9K

Likes

2.2K

Comments

144

Shares

150

SoundTikTok sound · Hold On, We're Going Home - Album Version

View velocity

+5.8K views over 9.9d · ~583/day · 5 observations

Promoted product

Hermes Agent

Hermes Agent — a standalone terminal app and a native application for macOS, Windows, and Linux. Install it and start a conversation with the agent that grows with you.

15 videos13 creators

Why it worked

The video taps into the growing interest in local AI models and provides a practical demonstration of running and comparing different models, appealing to a niche but engaged audience interested in AI hardware and software.

Summary

The creator showcases their current AI stack, highlighting the use of SuperDeepseek V4 and Qwen3.8 models. They discuss running these models locally using vLLM and compare their performance to Opus 5, expressing satisfaction with the local AI capabilities.

Structure

1Introduction to current AI stack
2Showcasing hardware and models (SuperDeepseek, Qwen)
3Discussing model performance and comparison (vLLM, Opus 5)
4

Original caption

Hardware is getting more capable by the week thanks to the software! #localai #qwen #deepseek

More from @tys.ais

Product placement

SuperDeepseek V4 Flash Abliterated (ACTS - it is a model being run and benchmarked, removing it would change the video's subject) Qwen3.8 27B Unsloth Nvfp4 (ACTS - it is a model being run and benchmarked, removing it would change the video's subject) Qwen (APPEARS - it is the account that posted the tweet, removing it would not change what happens in the video) SGLang (APPEARS - it is the account that posted the tweet, removing it would not change what happens in the video) vLLM (ACTS - it is a tool being used to run models, removing it would change the video's subject) Opus 5 (ACTS - it is a model being compared against, removing it would change the video's subject) Hermes Agent (ACTS - it is a UI for running models, removing it would change the video's subject)

On-screen text

Current AI stack Run Qwen3.8 27B locally Not on Opus 5 level* Even though Opus 5 has been garbage lately HERMES AGENT Using mtp for now*

Transcript

oh yes. running flash 00:04 0753 two across two of these sparks that you see here one running 00:09 3.8. getting second, and then the qwen is at thirty. The Qwen model dropped today and like I don't like running lang. I like vLLM more not option for 3.8 yet, but when there is is Using mtp for now* it. I've been extremely impressed with Deepseek even or codex anymore. my is hermes super and 3.8. models context. to the on they and feel just level now locally at unlimited have them working as mixture of agents individually. should Let me know what I should build.

I’m pretty heavy on owning your weights and keeping your data private but if you

DeepSeek-V4-Flash-DSPark