About this video
Your base MacBook is lying to you about its AI capabilities. Xiaomi claims their 9b Mimo model makes every 16gb Mac a free coding agent, but my real world tests tell a different story. Key Takeaways: - Mimo 9b achieved an 83% success rate on real world Next.js tasks. - Performance is 10 times slower than Claude Sonnet. - Memory usage spikes to 20gb in long sessions, risking system crashes on 16gb machines. - Specific configuration changes are required to move from a 6% to an 83% success rate. - It struggles with specific library implementations like Zod compared to larger models.
Your Base MacBook Is Not Ready For The Local AI Revolution
Xiaomi claims their new 9b parameter model, Mimo, is the holy grail for local development. They say it scores 61% on Sweetbench and runs perfectly on a 16gb Mac. I put those claims to the test on my own Next.js site using 10 real world tasks to see if your hardware can actually handle the heat.
The Benchmark Versus Reality
While benchmarks like Sweetbench are useful, they rarely reflect the daily grind of a software engineer. I gave Mimo 10 specific jobs on my website, including refactoring content forms across 3 files and building a rate limiter from a specification.
With the right settings, the model achieved an 83% success rate, scoring 25 out of 30 across 3 attempts per task. It even beat its own benchmark of 61%. However, getting to that 83% required significant tweaking.
The 16gb Memory Trap
The marketing says it fits in 16gb, and technically, it does: at least when it first loads. The model occupies about 9.5gb of memory initially. However, during long coding sessions, I observed the memory usage creeping up to 18gb, 19gb, and even 20gb.
On a 16gb Mac, this is a disaster. Between macOS, your browser, and your IDE, there is no room left for a model that refuses to stay within its budget. You will likely face crashes or a system that slows to a crawl.
How To Optimise Your Setup
If you want to try this yourself, do not use the default settings. Here are the 3 changes that moved my score from 2 out of 30 to 25 out of 30:
- Context Window: Increase it to 64,000 tokens.
- Reply Cap: Limit the model's reply window to 8,000 tokens to save memory.
- Avoid Lean Flags: Do not use 'bare' flags that hide essential file creation tools.
The Verdict
Mimo is a powerful local model, but it is not a Claude Sonnet killer yet. Sonnet is 10 times faster and far more accurate with complex library logic. If you have a 16gb Mac, use Mimo for short, isolated fixes, but do not expect it to replace your cloud subscriptions just yet.
Transcript▾
Xiaomi says this 9b parameter model scores 61% on Sweetbench, the test where AI fixes real bugs in real projects. Technically, this can run on a base 16gb Mac. So I gave it 10 jobs on my own website. That is it refactoring my content form across 3 files. The tests passed and nothing went to the cloud. It is a real coding model, but the 16gb Mac part cannot sign that off. If this is true, every base MacBook is a free coding agent, and that is a big deal. By the end, you will know how it scored on real code and how it compares with Claude Sonnet and the settings you need to make it work. If you are on an 8gb Mac, this 1 is not for you. And there is 1 number in my results that goes straight at the 16gb claim, and I will show you that near the end.
Let us start with the easy bit. Does it fit? First, what you are actually downloading. The name says Mimo, but this is not Xiaomi's big model. It is that model distilled into a 9b parameter Quen, and it comes in at 8.7gb. I tested it on my M5 Max, so I capped the graphics memory at what a 16gb Mac gets by default. That is about 10.7gb. Memory carries over between Macs, but speed does not. At 32,000 tokens of context, it used 9.8gb. Fits. At 64,000, it is 10.9, right on that line, but it is still loaded under my cap. On my Mac, it ran 50 to 80 tokens a second. On a base 16 Mac, I would guess 15 to 25. That is my estimate, not actually a measurement. So on paper, yeah, it does fit.
Sweetbench is a benchmark, and that 61 is Xiaomi's own number. But I wanted to know how it does on code it has never seen. So I wrote 10 tasks on my Next.js website. Fix how URLs handle accented letters, show dates the British way, close a security hole, build a rate limiter from spec, split a forms validation into its own file. Each 1 has tests that fail until the job is done. That is TDD. The grader runs my original tests, so it cannot cheat by editing them, and it never tried. 3 attempts per task, so that is 30 in total with Claude Code's agent running the local model through Ollama. Best result, 25 out of 30, and that is 83%. Every easy task and most of the hard ones, including a 3 file refactor and the rate limiter. So on my code it actually beat its own benchmark.
If you want more tests like this on real code and not just benchmarks, then please subscribe. Here is the bit that nobody puts on the model card. My first run with default settings scored 2 out of 30, and if I stopped there this video would say the model is rubbish. It was not the model. 3 things were wrong. 1, Claude Code keeps 32,000 tokens free for the model's reply in a small window that leaves no room to work. Cap that at 8,000. 2, give it 64,000 tokens of context, not 32,000. That is still inside the 16gb budget just. 3, I had used a flag called bare to keep Claude Code lean. It quietly hides the tools for creating files and searching code. Just do not use it. Same model and the same tasks, 2 out of 30 went to 25. Should I try a lighter agent like Pi or Open Code next? Let me know down in the comments.
For a reference point, I ran Claude Sonnet on the same 10 tasks, 30 out of 30 with a median of 21 seconds a task. Mimo's median was about 201 seconds, so it is about 10 times slower and misses 1 task in 6. I also squeezed Sonnet into the same small 32,000 window. It still got 30 out of 30. Mimo at 32,000 got 12, so the small window does not explain everything. The model needs more room to think than Sonnet does. 1 more weakness: a task needed a specific way of reporting errors in Zod, the validation library. Sonnet used it every time. Mimo never did in 21 tries, and 2 of its passing answers left TypeScript errors behind. You would still want to review its work.
Now the number I promised. The model loads at about 9.5gb. Short tasks, a couple of minutes, stay at 10 to 12gb. That is fine. But long sessions kept growing to 18, 19, and nearly 20gb every time, even with my 16gb cap in place. Because that cap only controls how the model is set up when it loads, it does not stop memory growing during a session. A real 16gb Mac has 16gb for everything: macOS, your browser, your editor. So 1 of 3 things happen. It frees the memory and carries on. It slows to a crawl, or the model crashes. I cannot tell which 1 from my Mac. If you have got a 16gb, try it and let me know.
So the title, 9b. Yes, but it is a distill, not the full Mimo. 61% on Sweetbench. On my code, it did better: 83% but only once I fixed the settings. On a 16gb Mac, it fits when it loads. In long sessions, I do not think it stays there. Like I said at the start, it is a real coding model. The 16gb part is what I cannot sign off. Would I use it on a 16gb Mac? For small, clear fixes in a short session, yes. As a Sonnet replacement, no. If you want to set up a local model with Claude Code from scratch, that is this video.