Is Qwen 3.8 27b The New King of the Local LLMs? — Transcript
Full transcript
- 0:00So, Qwen 3.8 27-billion parameter local
- 0:03AI model has been out for a few days
- 0:05now, and I've been testing it to see
- 0:07what it is capable of. And I've been
- 0:09surprised on how well it is performing,
- 0:11especially compared to the models that
- 0:13we had what, just a few weeks ago. This
- 0:16is another major step change in what you
- 0:19can achieve locally on local hardware
- 0:22without having to send everything up
- 0:24into the cloud. That's the topic of
- 0:25today's video. So, if you want to find
- 0:27out more, please let me explain.
- 0:32So, Qwen 3.8 is a 27-billion parameter
- 0:34model. Now, at its full size at 16-bit
- 0:38floating-point numbers, it's a 55-GB
- 0:41download, and you're probably going to
- 0:42need 128-GB of RAM to run it, which is
- 0:46beyond what most people have as VRAM,
- 0:49beyond what most people have kind of at
- 0:51home, even with a very good gaming
- 0:53setup. However, if you do have something
- 0:56like an RTX 5090 or something else with
- 0:5832-GB of VRAM, then the
- 1:02>> [snorts]
- 1:02>> 4-bit quantized version of Qwen 3.8
- 1:0527-billion is runnable on your local PC,
- 1:09and actually at a fairly decent rate. I
- 1:11was getting 96 tokens a second running
- 1:15it on an RTX 5090. That means you can
- 1:17run it locally on your PC and do
- 1:19everything you'd want to do with a
- 1:21normal large language model out in the
- 1:23cloud, but now you're running it
- 1:24locally. And I've been testing it
- 1:26specifically for how good is it at you
- 1:28doing coding using open code as the
- 1:32coding harness. But first, let's start
- 1:34with a few simple things. Does it answer
- 1:36the Alice question? Yes, it does, no
- 1:39problem, as you would expect. Does it
- 1:41handle the 7-minute egg timer and the
- 1:4411-minute egg timer question? Yes, it
- 1:47does, no problem. So, now those are out
- 1:50of the way, we can kind of get a feel
- 1:51for the kind of capabilities of this
- 1:53model, let's push it a a further. The
- 1:56first thing I did, cuz I wanted to have
- 1:57a bit of fun, was to get it to create a
- 1:59Space Invaders game using a single HTML
- 2:02file using inline JavaScript and no
- 2:05external libraries. So, this is just had
- 2:08to create everything itself inside the
- 2:10HTML file. It did that. It produced me
- 2:13index.html. You can double-click it and
- 2:15you get a nice Space Invaders game. So,
- 2:18I wanted to push it a bit more. So, I
- 2:19used my protocol encoding type length
- 2:22value encoding question. I asked it to
- 2:24create a specification similar to ASN.1,
- 2:28similar to Google's Protobuf. Now, it
- 2:30created that specification no trouble.
- 2:32There were a few tiny mistakes in it. It
- 2:34did all the specification defining how
- 2:36the protocol should be laid out. But,
- 2:38then at the very end it gives some
- 2:39working examples. This is what happens
- 2:41if you want to encode an integer. This
- 2:42is what happens if you want to encode a
- 2:43string. And actually one of those or two
- 2:46of those had a minor error in the
- 2:48numbers it was telling you that you
- 2:50should use. So, the examples it was
- 2:51giving uh didn't actually quite work
- 2:53out. But, if you fix those, no problem
- 2:55whatsoever. So, I then went on to say,
- 2:57"Well, now implement it including
- 2:58testing, including end-to-end testing,
- 3:00unit testing, and so on." And I did open
- 3:03code. Let it run. A few times I had to
- 3:05interact with it telling it to, you
- 3:07know, "Here's a plan. This is what you
- 3:09should do. Now, go ahead and do it." And
- 3:11just as you would do with normal coding
- 3:14with a harness. And it worked perfectly.
- 3:16It gave me a nice little library that
- 3:19could implement the specification
- 3:20including some end-to-end testing,
- 3:21including some unit tests. Worked
- 3:23absolutely fine. So, my big test after
- 3:26that, as I've done in several other
- 3:27videos here on this channel, is my new
- 3:30Scrippy uh programming language. So, the
- 3:32idea is I just give it a example of the
- 3:35of the new Scrippy language. I didn't
- 3:37define it formally. I tell it a little
- 3:39bit about it like it doesn't use
- 3:41semicolons. I tell it it's an
- 3:43interpreter. I tell it it's written in
- 3:45C. And then I say, "Go ahead. Knock
- 3:47yourself out. Go and write the
- 3:49interpreter for this language." Now,
- 3:51that's quite a tall order, because
- 3:52there's no formal specification for the
- 3:54language, just an example code, and a
- 3:57lot of the stuff it had to kind of infer
- 3:59from the example code, and infer from
- 4:01what I've said in the bit of text. Now,
- 4:04I've tried this before in other videos,
- 4:05and really local models weren't able to
- 4:08do it even Qwen 3.6. They weren't able
- 4:11to do it. They kind of either got stuck
- 4:13in a loop, just not being able to finish
- 4:15it, or what they produced crashed all
- 4:17the time, had to severe memory leaks. It
- 4:19was a bit of a disaster. However, the
- 4:21amazing news is here is that in a
- 4:24relatively short amount of time,
- 4:26certainly wasn't even overnight, just a
- 4:28few hours, again, few promptings, there
- 4:31was a plan, and then we went through the
- 4:33different steps on how to implement it,
- 4:35and then, you know, I said, "Also write
- 4:37some documentation." And I told it all
- 4:38the stuff back and forth as you would
- 4:40with a coding harness. It worked.
- 4:42Absolutely worked. You can run new
- 4:44Scrippy. You It created its own set of
- 4:47test scripts to run as well. Uh and the
- 4:51example script that I gave it for the
- 4:52readme file, that runs. And so, there
- 4:55you go. This local LLM running on a PC,
- 5:00Qwen 3.8, was able to do all of that
- 5:02coding using open code. You could use
- 5:04other harnesses as well, I'm sure.
- 5:06Haven't tested them specifically, but
- 5:07I'm sure it'll be just as good. And it
- 5:09did it all for me. And it actually
- 5:10works. So, this is really amazing that
- 5:12we've reached this point. Uh and I'm
- 5:14sure things are going to get better. But
- 5:17this is amazing that we've reached this
- 5:18point for local LLMs, just the power and
- 5:21the capabilities that they have. Okay,
- 5:23I'd love to hear your thoughts in the
- 5:24comments. Have you tried Qwen 3.8? Do
- 5:27you use a local AI for coding, or do you
- 5:31only rely on the big models? Do you not
- 5:33do any coding on it? Love to hear your
- 5:35thoughts in the comments below. Okay,
- 5:37that's it. I'll see you in the next one.
- 5:43>> [music]
- 5:58>> Ah.
About this transcript
This page contains the full transcript of Is Qwen 3.8 27b The New King of the Local LLMs? by Gary Explains, generated from the public captions YouTube serves with the video. The transcript has 1,101 words across 153 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.