The DeepSeek Effect: How Open-Source AI Got Good Enough to Run at Home

What DeepSeek Actually Did

Before DeepSeek, the AI industry had a simple assumption: bigger models need bigger computers. Want a smarter AI? Buy more GPUs. Want the smartest AI? Build a data centre. This is why ChatGPT costs billions to run and why you pay $20/month to access it.

DeepSeek challenged that assumption with a technique called Mixture of Experts (MoE).

Here's the idea: instead of using every part of the model for every question, MoE divides the model into specialised "expert" modules. When you ask a question, the model activates only the 2-3 experts that are relevant to your request. The rest stay dormant.

The result? A model with 236 billion total parameters that only activates 21 billion at any given time. It has the knowledge of a massive model but the speed and resource requirements of a small one. Same intelligence, fraction of the cost.

The Ripple Effect

DeepSeek didn't just build a clever model — they proved that the "bigger is better" approach was wasteful. And because they open-sourced their work, every AI lab in the world could learn from their techniques.

What followed was an acceleration in open-source AI efficiency that nobody predicted:

• Models got smaller — techniques like quantisation and distillation compressed models from hundreds of gigabytes to single-digit gigabytes without catastrophic quality loss
• Models got smarter — MoE and architectural improvements meant that a 7B parameter model in 2026 is smarter than a 70B model from 2024
• Hardware requirements dropped — what once needed a server-grade GPU (A100, H100) now runs on a consumer RTX 4060 with 8 GB of VRAM
• Speed improved — smaller active parameter counts meant faster responses, often matching or exceeding cloud API latency

Big Tech vs. Open Source: Two Different Races

The AI industry is quietly splitting into two worlds.

Big tech — OpenAI, Google, Anthropic — is chasing AGI (Artificial General Intelligence). They're building trillion-parameter models, spending billions on compute, and funding it all with your monthly subscription. Their pitch: pay us $20-200/month and we'll give you the smartest AI on the planet. The trade-off? Your data flows through their servers, your access depends on their terms, and prices only go up.

Open source — Meta, Mistral, DeepSeek, Alibaba, and thousands of independent researchers — is solving a different problem: how good can we make AI that runs on hardware people already own? Their models are free, run offline, and keep your data completely private. They're not trying to build a god — they're trying to build a tool that works.

And here's the key insight: for most everyday tasks, the open-source tools are already good enough. Transcription, writing assistance, code generation, translation, summarisation — local models handle all of these competently. You only need cloud AI for the most complex frontier tasks, and even that gap closes a little more each month.

What Happens When the Bubble Bursts?

The cloud AI industry runs on venture capital, subscription revenue, and the promise of future capabilities. When the economics tighten — when investors want returns, when subscribers push back on price increases, when the next big capability leap doesn't arrive on schedule — things get interesting.

Local AI doesn't have this problem. The model on your hard drive will work the same in five years as it does today. It doesn't need a subscription, a server, or a company to keep existing. It's a file on your computer. Nobody can revoke it, reprice it, or retire it.

This is why the transition to private, local AI feels inevitable. Not because cloud AI is bad — it's excellent — but because owning your tools will always beat renting them, the moment the owned version is good enough.

We're at that moment now. For an increasing number of tasks, it's good enough.

What This Means for You

If you have a PC with a modern GPU — the kind used for gaming, content creation, or even just a decent work laptop — you can run AI models that would have cost thousands per month in cloud compute just two years ago. Tools like Ollama make it trivially easy to get started.

For transcription specifically, this is where Vox Bar lives. The Voxtral model — built by Mistral, the same lab that helped drive the efficiency revolution — runs entirely on your GPU. Your voice stays on your machine. Your text stays on your machine. The AI that processes it is a file on your hard drive that nobody else can access, update, or monitor.

While big tech debates the ethics of training on your data, local AI has a simpler answer: your data never leaves your room.