The latest generation of open-weight large language models have been absurd. We’re now getting what people have started calling “Mythos-class” models: enormous, extremely capable models that are increasingly difficult to distinguish from the frontier stuff in actual use.
Moonshot AI dropped its latest Kimi models with fantastic benchmark results (so good that Anthropic panicked and gave everyone $100 of free Fable credit), then came DeepSeek with V4 Pro, and Alibaba with Qwen 3.8.
Qwen 3.8 is the one I’m most excited about, though, because unlike the other new Mythos-class models, it comes with an official 27B weight. That’s small enough to comfortably fit on a consumer GPU like an RTX 3090.
I’ve said time and time again that local language models are no longer just an “oh, that’s cool” thing. They’re getting good and efficient enough that you can actually delegate real work to them. In my previous HomeLab entry, I said I mostly used my local AI setup for easy but high-volume tasks, usually with strict JSON output.
Well, that just changed, my friends. Qwen 3.8 27B is mind-blowing.
I put Qwen 3.8 against Claude and GPT
Big weight gap, but the results…
The first thing I do when I get my hands on a new model nowadays is the pelican test. I’ve written about using the pelican test to benchmark LLMs before, and you can see a catalog of results in that article.
So, I threw the same simple prompt at Qwen 3.8. You can see the result above. Isn’t that absolutely mind-boggling? Here I was, ready to justify all the shortcomings because this is a much smaller model and so on. I can’t even find the shortcomings to begin with. What’s going on? Just compare it to the other models below. Qwen 3.8 is better than Opus 5! Of course, that’s just for this test. Maybe Alibaba knew and trained Qwen 3.8 specifically to be better at drawing SVG pelicans.
Qwen 3.8 27B is a dense model. That means all 27B parameters are active. Despite that, and despite the magnificent pelican it drew, it’s also very fast. Even on my measly RTX 3090, I got this result at 75 tokens per second, which is really fast.
Then I asked it to design an entire blog
Next, I did the boring test: ask it to build a blog. It was, once again, fantastic. It looked great! It’s got great taste, and it did a thorough job. Different pages, sensible stubs, everything beautiful and functional.
It does look like it’s taking some design cues from Anthropic’s book, but hey, it looks great!
I feel compelled to constantly remind you that this is a 27B model running on my own machine. Fully local.
Then Qwen (almost) nailed my hardest coding test
I really expected it to fail here
The task is to build a full fluid physics simulator. The constraint is that the agent can’t just use an available library and slap a UI on it. It has to build the simulator from scratch.
Notice that I said agent rather than model. This is a rather ambitious feat, and it only makes sense to have a harness for it. That’s what I tested Fable (Claude Code) and GPT 5.6 Sol (Codex) with, so I needed a harness for Qwen 3.8. The harness I picked was DeepSeek Harness (I’ll talk more about it in a future article).
You can see Fable and GPT’s results in my previous article comparing them. This test is my favorite, but it’s also the most expensive. I wanted to use it as a failure example for Qwen 3.8, but… Qwen didn’t fail. It nailed it.
Qwen took nearly three hours to output this, and that was with my intervention to cut it short. At some point during its performance testing, I told it that it was taking too long and should move on with its best solution instead of continuing to test. That got me the result you see now.
The only real flaw is performance. Above 1K particles, the simulator starts dropping frames, and above 3K it’s unusable. That’s still quite something, given that most other agents just put a hard cap of 2K particles on their simulators.
I can’t emphasize this enough. In the same article, you can see that Grok Build failed. Now, here I have a local model, using a local harness, running entirely on my own machine, that delivers where some frontier models fail.
My local Qwen is tiny compared to the frontier models
And somehow, we’re still having this conversation
OK. A quick clarification. This is not a fair comparison. If I wanted to make it fair, I’d have to use the full-size Qwen 3.8 model. I’m not. Not only that, I’m running Qwen on my own hardware.
27B is the smallest model in the Qwen 3.8 family. The other, full-size one, is the gigantic Qwen3.8-2.4T-A95B. That’s 2.4 trillion parameters.
The VRAM requirement math can be simplified to (parameters × bits per weight / 8) + the KV cache + additional capabilities or mini models (e.g. vision).
In my case, I used the Q4_K_M quant. For that, I needed 17.8GB for the model alone, and with 100K of context on a Q8 KV cache, it took up a total of 22GB of VRAM out of the 24GB I had to spare. Tight fit!
To put that into perspective, if I were to run the frontier Qwen 3.8 with 2.4 trillion parameters, even using the same Q4_K_M quant, I’d need around 1.5 terabytes of VRAM just to run that one model.
I say this to put the comparison into perspective. ChatGPT and Claude are that class of model. They are massive, trillion-parameter-scale systems. So, it’s my 24GB of VRAM versus their terabytes — and I didn’t fare so badly.
So what’s actually left for the big AI companies?
At this point, the main advantage the big cloud models have over my local setup is the context window. I can’t push much further than 100K tokens without breaking my card, whereas the latest frontier models provide context windows of 2 million tokens or more.
Context is super easy to burn, especially with a vocal thinker like Qwen 3.8, but there are tools to work around this. For example, you can have a much smaller model on the side compact the context and keep only what’s important, or systematically rip out old tool calls to free up space. But it’s not perfect.
The other major advantage of subscribing to the big AI models was their harnesses. Claude has Claude Code and Cowork. ChatGPT has Codex and Work. But that’s not as much of an advantage anymore, thanks to the growth of open-source harnesses like Qwen Code and DeepSeek Harness.
It’s pretty ironic. OpenAI, Anthropic, Meta, and Google will keep lecturing us that AI shouldn’t be a monopoly and that everyone should have access to high-quality models. I haven’t heard the same platitudes from Alibaba, and yet, here we are.
Qwen 3.8 is the high-quality local model we wanted. And it will only get better from here.