About this video
Local AI is no longer a hobbyist toy; it is a direct threat to the silicon giants. In this video, we deep dive into the brand new Qwen 3.8, a 27B dense model from Alibaba that is shaking up the local LLM rankings. We put it through a real-world coding stress test to see if it can actually replace your cloud-based subscriptions. Key Takeaways: - Qwen 3.8 Overview: Exploring the 27B parameter dense architecture and its 262k context window. - Mem Zero Integration: How to give your local AI a permanent memory for consistent coding sessions. - Performance Benchmarking: See how Qwen 3.8 stacks up against Google's Gemma and DeepSeek. - Real-World Coding: Watch as the model attempts to build a full OCR PDF scanning application from scratch. - Hardware Guide: What you actually need to run this model smoothly, from MacBook Airs to M5 Max monsters.
Alibaba Just Dropped a Model That Makes Your Cloud Subscription Look Like a Waste of Money
The gatekeepers of Silicon Valley are losing their grip on the crown of artificial intelligence. With the release of Qwen 3.8, a 27 billion parameter dense model, Alibaba has proven that high-level intelligence no longer requires a monthly subscription or a massive server farm. If you have been waiting for the moment local AI becomes a viable replacement for the industry leaders, this might be it.
The 27B Powerhouse
Qwen 3.8 is a dense model, which differentiates it from the Mixture of Experts (MoE) architectures we often see in coding-specific tools. Dense models are built for raw intelligence and deep research. This version boasts a default context length of 262,144 tokens, which can be extended to a staggering one million. During my testing on an M5 Max, it pushed through complex tasks with a surprising level of nuance, though it certainly demands respect from your hardware.
Memory is the Missing Link (SPONSORED)
One of the biggest hurdles in local AI is the lack of persistence. Every time you start a new session, your agent forgets who you are and how you like to work. This is where Mem Zero comes in. By providing a persistent memory layer, it allows your local models to remember specific preferences, such as coding conventions or previous corrections. It is the bridge between a one-off chat and a genuine digital assistant.
https://samuelgregory.co.uk/mem0
Putting Coding to the Test
I tasked Qwen 3.8 with building a local OCR application for scanning PDF documents. The results were a mixed bag. While the model was incredibly fast (hitting 21 tokens per second) and created a surprisingly refined UI, it struggled with the specific complexities of multimodal OCR logic. It spent over two hours refining a spec that a larger, cloud-based model might have handled more elegantly.
The takeaway is clear: dense models are brilliant for research and general intelligence, but they might not yet be the "coding killers" that specialised MoE models claim to be.
Hardware Requirements
You do not need a supercomputer to run this, but you will need more than a basic setup. While I tested this on an M5 Max with 120GB of RAM, the 4-bit quantised version should sit comfortably on a 32GB MacBook Air. If you are serious about local AI, prioritising RAM is no longer optional; it is a requirement.
Final Verdict
Qwen 3.8 is a valid improvement on the local AI landscape. It is smart, it is fast, and it is local. If your work involves heavy research and high-context analysis, it belongs in your toolkit. However, for agentic coding workflows, you might find more success elsewhere. Keeping it local has never looked this good.
Transcript▾
Qwen is one of the top contenders for running AI locally, alongside things like Google's Gemini, DeepSeek, Kimiko, and GLM. They just released their latest model 3.8, which is a 27 billion parameter dense model. In this video, we are going to put it through its paces and show what it is like to get real work done with this latest and greatest from Alibaba.
Before that, a question I get asked all the time is how to persist memory between coding sessions. That is where today's sponsor, Mem Zero, comes in. It is a plugin for Claude Code or Codex, and it helps maintain consistency across your workflow. For example, if I need to correct specific terms in my podcast captions, Mem Zero remembers those corrections for future sessions.
Looking at the model card, it boasts coding capabilities, but it is a dense model. These are typically better for research and intelligence rather than just coding, which is a specific field of expertise. One thing I like is that it is a vision model. It has a default 262,144 context length, extendable to a million tokens.
In benchmarks, it is performing very well against Gemma. I am using the 4-bit quantisation version, which fits into about 16GB of VRAM. You will still want at least 24GB for a decent context window. On my M5 Max with 120GB of RAM, we are going to try and build an OCR app.
The first prompt took 24 minutes to complete. It used a Python back end and HTML front end. The UI looked a bit messy initially, but after some feedback, it became quite refined. The OCR functionality was a bit jumbled at first, but with further prompting and a better spec, we got a working result.
Overall, Qwen 3 is a valid improvement in intelligence. While I might not use it for every coding task due to its dense nature, it is a fantastic contender for research, especially if you are running it on something like a 32GB MacBook Air.