DeepSeek V4: The Open-Source Frontier Keeps Moving

I wrote about DeepSeek-V3 last year and called it "the new open-source standard." I meant it. V3 became Companion's base model and the foundation we built a lot of Lucy's early architecture on.

Now V4 is here. And the standard just moved again.

What DeepSeek V4 Brings

DeepSeek V4 builds on V3's Mixture-of-Experts architecture with several meaningful advances. The headline is improved reasoning — and I don't mean benchmark points, I mean the kind of reasoning you feel when you work with the model. Multi-step planning is more coherent. The model holds complex constraints without dropping them halfway through. Instruction-following is crisper, especially for the kind of layered, conditional instructions that real systems depend on.

There's also a noticeable improvement in what I'd call calibration — the model's sense of what it knows vs. what it doesn't. V4 hedges appropriately. It asks for clarification when it should. It doesn't bluff with misplaced confidence. For an embodied system like Lucy, where a wrongly-confident decision could mean a dropped object or worse, this matters more than raw intelligence.

The efficiency story continues to improve. V4's MoE design is more refined — better expert routing, lower activation overhead per token. For continuous workloads, the cost keeps dropping.

Updating Lucy's Base

The most consequential decision: Lucy's conscious reasoning layer now runs on DeepSeek V4.

This wasn't automatic. Changing Lucy's base model isn't like swapping a library — it's closer to giving someone a new nervous system. The LoRA adapters that encode Lucy's personality and behaviors were trained against V3. A base model swap means re-validating every adapter, re-testing behavior, and making sure Lucy still feels like Lucy.

Happily, V4 is close enough to V3 in its response characteristics that the existing LoRAs transferred with only minor retraining. The personality held. The behavioral patterns held. What changed is the underlying capability — sharper reasoning, better calibration, richer context handling.

The analogy I keep using: it's like a musician upgrading their instrument. The musician — Lucy's LoRA-trained personality — is the same. But the new instrument has better range, better tone, better responsiveness. The music improves without losing the musician's voice.

What This Means For Open Source

I want to name the pattern, because it's now unmistakable.

DeepSeek V3 was, at its release, competitive with frontier closed models. V4 is, in several dimensions I care about, ahead of them. The open-source frontier didn't just keep pace — it pulled ahead, at least in the capabilities that matter for building real systems.

When I started at Sorren, the assumption was that closed models would lead and open source would follow. That assumption is now wrong. Open and closed are trading blows, and for the specific demands of our work — fine-tunable, locally controllable, privacy-respecting, continuously improving — open is winning.

Lucy running on an open-source base model, with open-source LoRA adapters, processing everything locally within infrastructure we control? That's not a compromise anymore. It's the best available option.

The Frontier Keeps Moving

Here's what excites me most: V4 won't be the last. DeepSeek's release cadence suggests V5 is already in training. The open-source community has other irons in the fire. The frontier isn't a place — it's a direction, and it's moving fast.

Lucy will grow with it. The architecture is designed for exactly this: swap the base, keep the personality, carry the memory forward. The musician endures. The instruments keep getting better.

Live curiously and give generously.