I tried replicating fully agentic coding on my laptop and I'm impressed! My setup - Qwen 27bn (4bit quant) - Pi coding agent - CCO sandboxing for yolo mode My Mac M3 has 36GB ram. It's surreal to see a local model read prompt, follow it flawlessly and then... build a fully self-contained html/css page with zero errors. I got about ~20 tokens/second, and on this prompt quality similar to Haiku! But... this is running locally :)show more

The page it generated from my prompt:

bhai, it can go upto ~55t/sec. this is on 16gb ram

Yeah - didn't Julien Chaumond post about a similar setup? This is a good fully local Agentic Coding harness for those with good Macs

yeah, got inspired by that post (+ locallama bois) and wanted to try it myself

Love seeing this, and also totally wild you’re doing this on an M3 w/36gb of ram. Also, how did you get 36gb? I thought it was only 32 or 48 that you could get?

My macbook is a bit old

I have M3 Pro 36GB too. With lm studio, I am barely hitting 10 tokens/s on same model. You should make a short video of your setup. Many YouTubers making 30mins video and end result is still slow. Btw what’s your context length and prompt processing time?

i use llama.cpp and this is the config llama-server --model ~/models/qwen3.6-27b/Qwen3.6-27B-UD-Q4_K_XL.gguf --mmproj ~/models/qwen3.6-27b/mmproj-BF16.gguf --alias qwen3.6-27b -c 131072 -n 32768 --no-context-shift --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.00 --presence-penalty 0.00 --fit on -fa on -ctk q8_0 -ctv q8_0 --chat-template-kwargs '{"preserve_thinking": true}'

that's cool! consider giving a shot for the sandboxing part - it spins up microVM powered sandboxes in less than 1 sec and also has checkpointing.

Video is not 1x. Which model variant for 20 t/s ? Couldn’t get > 13 on my m4 pro 48 GB ram.

Yeah but you should use AEP (Agent Element Protocol) to achieve deterministic control over your agent outputs:

Nice… and if you need local dictation as well then try out Voibe -

How were the results? I bake-tested Gemma and Qwen against Opus, and there is still a sizable gap between them and Opus, especially on more technical design tasks. But I would bet my future that very soon, they will be as good as Opus is today.

I’m equally impressed. I am achieving 90 tok/s using the Qwen3.6 35B A3B model.

Local agent loops are getting real fast. Sandbox is what makes yolo mode tolerable for me.

27b at 4bit on 36gb ram is exactly why local agent stacks are viable now. run the quantized model for bulk work, escalate to frontier only when it fails. this is the economics i built my m4 pro setup around.

Cool demo but those benchmarks are for the full-weight model, not 4-bit quant which tanks coding accuracy pretty hard. Local agentic coding sounds great until you spend more time debugging hallucinated imports and wrong APIs than you would've writing it yourself.

I usually rely on 5.5 xhigh or Opus 4.7 for the heavy lifting, but local Qwen is surprisingly capable. It's perfect for running yolo mode without burning through your API limits.

@paraschopra that's wild! didn't think my laptop could handle something like this. gotta try it out myself.

what are the real life use cases you apply for the model? I am also trying to use these models .

Replicating agentic coding on a laptop is one thing, but scaling it to thousands of users will trade off latency and memory usage for multi-tenancy and security.

Have you tried the gemm4:e4b

À noter : agentic coding 27b 4bit qui tient sur Mac M3 36GB confirme la pression baissière sur l'inférence cloud pour les workflows dev. Pi + cco sandboxing change l'équation make-vs-buy pour les tooling agents. Le scaling laws s'inverse côté distribution.

People hyperfixate on the slight perplexity hit of 4-bit quants, but the actual unlock is freeing up enough VRAM so your KV cache can stretch its legs during long agentic loops.

wow pretty decent.

Pi mono is awesome to run with local models.

Could you do one test - For a complex project, can it do a simple change. Or simple repeated job ? ex: Check all REST API calls and see if they have standardized handling ? It can be a great internal tooling if it works. Tired of rate limits in coding agents.

That’s awesome—agentic coding on a Mac with a quantized 27B is legit. How’s Pi’s “yolo” mode feeling with the CCO sandbox? If you’re comparing approaches/costs for coding APIs, I’ve found OpenStacks ( handy to sanity-check plans.

dope! did you use anything to harness it?

Local is really fun, but pushing everything to the cloud is the game changer

Running local models works only when your project isn’t going to go into production. They are more of junior concept artists that give you practical clarity as to what something is ought to look like.

ordered m5 pro - 64 gb for running models locally

i built a similar local coding agent and connected discord with it. now i can just text the discord bot and it makes required changes without me having to open the mac!

the pi-dev harness is absolutely goat. Minimalist and like a swissknife

Local agentic coding gets real once you have sandboxing plus a tight eval loop; the model is the easy part. Nice setup.

Is there a sense of how many more tokens per second you would get on a 128GB RAM MacBook?

Looks good for simpler tasks like these. But can local models actually work on much more complex and huge projects.

How long do you think it will take until models with the capability of today’s SOTA can run on device?