Qwen 3.8 is a game-changer

I've been putting Qwen 3.8 27B through it's paces since it's release yesterday, and I'm liking what I'm seeing. It is a signifcant leap over it's predecessor Qwen 3.6 and joins a crowded landscape of "capable enough" open weight models like the final Deepseek V4 Flash and Tencent's Hy3 that are all either

1) runnable at home or on-site at a business/lab with reasonable hardware

2) or a factor of 100x or more cheaper to use as a service online compared to the incumbents (with the exception of GPT Luna).

This is a game-changer for agentic development, and one that comes at an inopportune time for the two AI giants that are racing towards their IPO.

Technical prowess

I used pi coding agent to let it implement fairly complex improvements for vllm, a major LLM inference engine, with a very good outcome:

It spent four hours hammering away without human interaction at improving inference performance for the outdated AMD Instinct MI50 data center GPU for certain Qwen 3.5/3.6 models - a GPU I happen to have in my lab. I had low expectations that this would be complete by morning 0 that it would at least require feedback from me (hello, Claude Code) or need to be motivated/coerced to continue beyond a "natural stopping point" (oh, hello again, Claude).

Initially I stuck to my lessons learned from working with Qwen 3.6 over the months. I gave it strong guidance ("look at these sources here, here's an example, ...") and had other models do adversarial reviews of any plans.

I'm not sure that this is still needed though:

Qwen 3.8 goes from simple plan executor to planning master

I used Qwen 3.5 and even 3.6 as workhorses: executing plans that much larger models like Claude Opus or GLM 5.2 came up with, and after these plans had been refined and reviewed. These models were simply to weak to write complex plans on their own, or to add much when reviewing existing plans.

Qwen 3.8 is the opposite. Here's what Claude thinks of a plan review done by it.

Both Claude and DeepSeek v4 are surprised at the depth and precision of the review, and even express it:

Thinking long and slow

One thing that is exacerbated in Qwen 3.8 which was already quite prevalent in Qwen 3.6 is that the model likes to think thoroughly. Depending on your setup this can make development workflows quite slow. There are tricks you can do to limit it, The good news: 3.8 allows you to set reasoning effort now, and there's also this trick you can do if you're serving it with llama.cpp:

--reasoning-budget 4096
--reasoning-budget-message "Wait, I'm overthinking this. Let's answer now."

Personally, I don't mind a longer reasoning process if it leads to a better outcome. And on my modest lab rig I get around 750 tokens/s in prefill and average 75 tokens/s in token generation. Setup used:

• 2 RTX5070 Ti 16GB
• llama.cpp
• spec dec MTP with depth 3