DeepSeek-V3: The New Open-Source Standard

There's a moment in cooking — if you've made the same dish enough times — when you stop following the recipe and start feeling the dish. You know by sound when the oil is hot enough. You know by smell when the garlic is about to turn. The recipe was scaffolding; eventually, you don't need it anymore, because the understanding has moved from your head to your hands.

Migrating Companion to DeepSeek-V3 this week felt a little like that. I've been working with DeepSeek's models long enough now that the migration was less "following instructions" and more "the next natural step." And the results have been genuinely exciting.

What DeepSeek-V3 Brings

DeepSeek-V3 represents a significant leap over its predecessors. The headline feature is its Mixture-of-Experts (MoE) architecture — a design where the model contains many specialized sub-networks ("experts"), but only a small subset are activated for any given token. You get the capacity of a much larger model while only paying the inference cost of a smaller one.

For those of us running models in production — especially continuously, like Companion's subconscious — this matters enormously. MoE means faster inference, lower memory footprint, and dramatically reduced compute costs per token. The model is bigger where it needs to be and lean where it doesn't.

The quality improvements are real, too. DeepSeek-V3's reasoning, instruction-following, and code generation are competitive with — and in some benchmarks exceed — frontier closed models. The fact that this level of capability is available in open weights feels like a tectonic shift.

Migrating Companion

I've been gradually moving Companion's stack toward open-source and locally-controllable components (see my posts on running DeepSeek LoRAs and testing Llama 3.2 Vision). DeepSeek-V3 felt like the right moment to make the base model migration official.

The conscious model — the one users interact with directly — now runs on DeepSeek-V3. The migration was smoother than I expected. V3's improved instruction-following made it more compliant with Companion's memory integration patterns. Its reasoning quality lifted the extended-thinking routing path (where complex queries go to Claude 3.7). And the general conversational quality — warmth, coherence, nuance — improved noticeably.

The LoRA adapters I trained for personality? They still work, with minor adjustments. The custom voices I've been shaping transferred over, which means Companion's personality isn't locked to a specific base model. That flexibility is exactly what I was hoping for.

The Subconscious Gets Cheaper

But here's the part that made me actually grin at my desk: the MoE architecture makes Companion's subconscious process dramatically cheaper to run.

Remember the economics I described in March — the always-on background process running on Groq infrastructure, summarizing conversations and updating the vector database? That process runs on a fast, inexpensive model. DeepSeek-V3's MoE design means each token the subconscious processes activates fewer parameters. Lower compute per token. Lower cost per token. The same continuous monitoring, summarization, and memory indexing — now running at a fraction of the previous cost.

This isn't a marginal optimization. It changes the unit economics enough that scaling Companion to more users — each with their own accumulating memory store, each needing 24/7 subconscious processing — becomes meaningfully more feasible. The infrastructure I was excited about in March just got a force multiplier.

What This Means for Open Source

I want to step back and name what's happening in the broader landscape, because I think it's historic.

A year ago, the gap between open-source and frontier closed models felt wide and possibly widening. Closed labs had the compute, the data, and the head start. Open-source was playing catch-up.

Today, that gap has narrowed dramatically — and for many use cases, closed. DeepSeek-V3 isn't a "good for an open model" model. It's a genuinely excellent model that happens to be open. When the best available option for a production system is also the one you can run, modify, fine-tune, and control entirely yourself, the dynamics of the entire industry shift.

For Companion — built on the premise of intimate, long-term, personal memory — this shift is foundational. The ability to run a state-of-the-art model, fine-tuned to a specific personality, processing sensitive personal context, all within infrastructure the user controls? That's not a feature. That's the vision becoming viable.

The Next Natural Step

I cooked that pasta aglio e olio again last night — the one I've made so many times I no longer need the recipe. It was better than the first time I made it. Not because the technique changed, but because my understanding deepened. The same steps, executed with more feel.

DeepSeek-V3 is like that. The architecture evolves. The understanding deepens. And Companion gets a little closer to what it's becoming.

Live curiously and give generously.