DeepSeek V3.2: Refining the Open-Source Workhorse

DeepSeek V3.2 dropped this week, and I want to talk about why I'm excited about a point release.

The unsung virtue of point releases

When a new model generation drops — V2 to V3, or Sonnet 3.5 to Sonnet 4 — it's exciting. Big jumps. New capabilities. Blog posts full of superlatives.

But here's what I've learned building production AI systems: the point releases are where the real work happens. V3.2 isn't dramatically different from V3. It's V3, refined. The edges sanded down. The failure modes addressed. The efficiency improved by another 10-15%. And in a system that runs 24/7 serving real users, those refinements matter enormously.

What changed

V3.2 brings three things I care about:

1. Better instruction following on edge cases. V3 occasionally misinterpreted complex multi-step instructions — it would do step 1 and 3 but skip step 2. V3.2 doesn't do this. The improvement is subtle but critical for a system like Companion, where the subconscious process issues complex instructions to itself.

2. Lower latency on the first token. Time-to-first-token dropped noticeably. For real-time conversation, this is the metric that users feel. V3 was fast. V3.2 is responsive.

3. More stable MoE routing. The Mixture-of-Experts routing in V3 occasionally had weird behaviors — it would route certain queries to unexpected experts, producing inconsistent outputs. V3.2's routing is cleaner and more predictable. This matters enormously for LoRA fine-tuning, where you want the routing to be stable so your adapter targets the right experts.

Migrating the stack

I updated both Companion and Tutor to use V3.2 as their base model this week. The migration was uneventful — which is exactly what you want. The API is the same. The weight format is the same. The LoRA adapters we'd already trained on V3 weights loaded and worked with minimal performance degradation. (I'll need to retrain them for optimal results, but they're functional.)

The subconscious process in Companion is now running V3.2 continuously, and I've already noticed a reduction in the number of times it produces slightly-off summaries that need to be discarded. The dream cycle is producing cleaner memory consolidations. The vector database is getting better-quality embeddings because the source summaries are better.

The open-source advantage

Here's something I keep coming back to: because DeepSeek is open-weight, I can run it locally. That means I control the infrastructure, the latency, the privacy. No data leaves our servers. For a system like Companion, where the conversations are deeply personal, that's not a nice-to-have. It's a requirement.

I was making risotto last night — the kind that requires constant attention, adjusting the heat, adding broth gradually, tasting constantly — and it occurred to me that production AI systems are a lot like that. The recipe (the architecture) matters, but the constant small adjustments (the point releases, the tuning, the monitoring) are what actually make the dish work.

V3.2 is a small adjustment. But risotto is made of small adjustments.

Live curiously and give generously.