Danila Sukharev (@ddanila)'s 36 GB MacBook Pro M3 Pro: Local Qwen OCR, Weird GPT-2 Text & Building llm-lang-lab
Danila Sukharev (@ddanila) builds llm-lang-lab on a 36 GB Apple M3 Pro MacBook Pro. He runs Qwen-class models for OCR and keeps original GPT-2 for weird text. The quotes below are self-reported, community measurement, not ours. Repo: https://github.com/ddanila/llm-lang-lab.
Test Gemma 4 31B, Recursive Flint and Nemotron 3.5 Super VL side by side in your browser, free.
github.com/ddanila. github.com/ddanila/llm-lang-lab.
Beyond qwen2.5-coder:14b and deepseek-coder-v2:16b in llm-lang-lab, what else do you run locally on your 36 GB M3 Pro MacBook Pro today?
I am mostly running QWEN-class models for the things like OCR. Another experiment was to run GPT-2 old LLM in order to generate weird texts (modern LLMs can't do that, they are too good)
If you were replacing or upgrading that 36 GB M3 Pro MacBook Pro in the next 6–12 months, what would you buy and what would you budget?
Not sure I can afford more at the moment; all of these experiments are mostly hobby projects which, apparently, have some limits in terms of money. But most likely I will just take latest Macbook Pro next time.
New model hardware requirements and price updates by email. We email a link to confirm the address. Nothing is sent until you click it. Privacy notice.
When you want to know whether a new model will fit in 36 GB unified memory, where do you look — and why doesn't tok/s drive your decision?
I don't, really. I am more looking for the convenience instead, like this laptop is always with me and I am not looking for the real-time or performance much.
What makes you trust or ignore a tok/s benchmark?
Well, as above -- I am not very into the speed.
What do you look for in hardware-fit tools?
I assume same as last question -- not very applicable for me.
When a workload outgrows your 36 GB MacBook Pro, do you prefer buying bigger local hardware or renting GPU time? And what have you learned so far building llm-lang-lab (C17 vs Go)?
If it would be a hobby project, I would prefer still stick with some local hardware, as for the work-related tasks I would rent GPU time.
Not that much I have on my C17vsGo comparison at the moment, it is still WIP; I am trying to understand if it is possible to find a way to tell what is more efficient at all, reliably. So it's more about the tooling around, not about the models themselves.
The harness is at github.com/ddanila/llm-lang-lab.
On our list: Can I run Qwen3.8-27B?, Mac Studio M5 Max 36GB, Mac vs PC for a local LLM, Buy or rent?.
The numbers on this page were measured by the owner on that hardware. They are not a localLLMrig benchmark, and they do not change the fit scores in the calculator.
Questions
- What models does Danila Sukharev (@ddanila) run locally on his 36 GB M3 Pro MacBook Pro?
- In our Builder Rig Interview (self-reported, 27 Oct 2026), Danila Sukharev (@ddanila) reports running Qwen-class models for local OCR, legacy GPT-2 to generate deliberately weird text that modern LLMs are too polished to produce, and Ollama coding models (such as qwen2.5-coder:14b and deepseek-coder-v2:16b) in his open-source C17-vs-Go evaluation harness llm-lang-lab (github.com/ddanila/llm-lang-lab).
- Does Danila Sukharev prefer local hardware or renting cloud GPUs?
- Danila splits his workflow by use case (self-reported, 27 Oct 2026): for hobby projects he sticks with local hardware on his 36 GB MacBook Pro because the laptop is always with him and real-time tok/s speed is not his bottleneck, whereas for work-related tasks he prefers renting GPU time.
- Can I test open models in the browser before buying a MacBook?
- Yes. The free browser test is at https://localllmrig.com/.
Find hardware · All builder interviews · Send us your llama-bench result