Student Machines

We are a design and research lab building natural interfaces that make complex AI workflows simple and affordable for everyone. Our flagship project is a friendly offline AI assistant named Modu.

Building the OS for local AI

We’re building the interface layer for an on‑device future, where people interact with local AI through clear, intuitive design. Today, most local AI interfaces are built for developers first. We design and test new UX systems that make deploying local AI frictionless for everyday users, bundled with the right tools and models out of the box.

Relevant metrics

2022

2023

2024

2025

2026

Hugging Face models

120k

200k

2.3m

2.8m

3.8m

AI PC shipments

21.5m

43m

114m

150m

200m

RAM required (GPT-4o)

110gb

45gb

20gb

10gb

9gb

Download size (GPT-4o)

140gb

45gb

50m

55m

60m

LLM users in Asia

180m

220m

256m

295m

340m

Hover on metric for details

Relevant metrics

2022

2023

2024

2025

2026

Hugging Face models

120k

200k

2.3m

2.8m

3.8m

AI PC shipments

21.5m

43m

114m

150m

200m

RAM required (GPT-4o)

110gb

45gb

20gb

10gb

9gb

Download size (GPT-4o)

140gb

45gb

50m

55m

60m

LLM users in Asia

180m

220m

256m

295m

340m

Hover on metric for details

While local models are becoming leaner and more sophisticated, public awareness of them remains low. The truth is, their real accessibility hinges on how seamless the onboarding experience feels. Today, there’s still a significant amount of friction in environment setup: you have to choose between thousands of constantly shifting model variants and parameter sizes, wire up the right MCP servers or API keys, know how to pip-install and run headless models, and even then, most local AI apps still collapse everything into a bare chatbox or CLI.


On their own, local models still fall short of what people now expect from everyday AI tools. It is when you layer on web search, multimodality, effectively unlimited file context, and rich MCP connections that they become viable for real, day‑to‑day workflows.

Top local models

<24b

Download

GPT-4o parity

Runs well on

12gb

96% (MMLU)

M2 Pro (16GB)

gpt-oss-20b

Open AI

5gb

85% (MMLU)

M1 Pro (16GB)

Qwen-3-8b

Alibaba

2.6gb

77% (MMLU)

M1 (8GB)

Gemma-3-4b

Google

2.5gb

74% (MMLU)

M1 (8GB)

Qwen-3-4b

Alibaba

Reasoning

Coding

Visual

Mixed

GPT 4o parity: localmodel MMLU / GPT 4o MMLU (88.7%) * 100

Top local models

<24b

Download

GPT-4o parity

Runs well on

12gb

96% (MMLU)

M2 Pro (16GB)

gpt-oss-20b

Open AI

5gb

85% (MMLU)

M1 Pro (16GB)

Qwen-3-8b

Alibaba

2.6gb

77% (MMLU)

M1 (8GB)

Gemma-3-4b

Google

2.5gb

74% (MMLU)

M1 (8GB)

Qwen-3-4b

Alibaba

Reasoning

Coding

Visual

Mixed

GPT 4o parity: localmodel MMLU / GPT 4o MMLU (88.7%) * 100

With the right setup, the advantages of local AI are immediately clear: ownership and control, privacy by default on-device, offline availability, and effectively zero marginal cost. Cloud models aren’t going anywhere, but everyone benefits from having a private, offline assistant as a dependable fallback: on the road with spotty Wi‑Fi, while working with sensitive files you don’t want leaving your device, or anytime you want capable AI that stays affordable over time. Because the truth is, for 80~90% of general AI tasks like summarization, analysis, drafting, and research, local models are more than capable.

Our building blocks

The thesis is that this is largely a design challenge: local AI will only feel compelling when it’s pre-packaged as a complete, everyday product, built around real workflows and instant usability, without complex configuration or manual setup. A truly accessible local AI application must have the following characteristics:

1

personalized onboarding

Downloads a model mix that fits your VRAM and use cases: lightweight model for quick responses, 20B for deep reasoning, vision model for OCR.

1

personalized onboarding

Downloads a model mix that fits your VRAM and use cases: lightweight model for quick responses, 20B for deep reasoning, vision model for OCR.

1

personalized onboarding

Downloads a model mix that fits your VRAM and use cases: lightweight model for quick responses, 20B for deep reasoning, vision model for OCR.

Student Machine Lab is run by Joonseo.

© 2026 Student Machine Lab

Built with ramen and kopi

© 2026 Student Machine Lab

Built with ramen and kopi