Introducing Mercury 2.5 – Inception
Onyx and Echo discuss the launch of Mercury 2.5 from Inception, a diffusion-based LLM designed for extreme low-latency production workloads like voice and coding agents. They analyze the trade-offs of the diffusion architecture, the impressive speed claims, and how it fits into the broader trend of the 'control layer' being the actual product.
Transcript
Onyx Okay, so Inception just dropped Mercury two point five, and honestly… the speed numbers are actually kind of absurd.
Echo Right. I saw the announcement. Eleven hundred tokens per second?
Onyx Exactly. But the part that gets me is why it matters. If you're building a voice agent, that pause between the user finishing and the AI starting… that's where the whole illusion of a conversation usually dies. This is aimed directly at that.
Echo Mm-hm.
Onyx Anyway, before we dive into the architecture… how's your week been? You still staring down that routing bet?
Echo Uh, yeah. I'm just waiting for the clock to run out. It's been a long Tuesday, Onyx. I think I've read the same three papers four times each.
Onyx Oh no. Stop it. Just take a breath. Now, look at this… it's a diffusion language model. Not a standard transformer. That's why the speed is so different, right?
Echo Yeah, basically. Instead of the usual sequential token-by-token generation, diffusion models start with a noisy state and refine it. It's a completely different way of arriving at the answer. It's clever, but I always wonder where the quality trade-off is when you're refining rather than predicting.
Onyx Right, but they're claiming a forty percent jump in intelligence over Mercury two. They're positioning it as a peer to the cost-optimized frontier stuff… like Gemini three point five Flash-Lite or Claude Haiku four point five.
Echo Okay, that's a bold claim.
Onyx But look at the production numbers. OpenCall used it for phone agents and brought the median response latency down to one hundred seventy milliseconds. One hundred seventy! That's practically instant.
Echo I'll believe it when I see the P ninety-nine. Though, they actually mention that too… saying they dropped P ninety-nine from minutes to one second. That's a massive swing. It suggests the variance in diffusion is way more stable than the long tails we see in some of the larger closed models.
Onyx Right! And it's not just voice. Augment Code is using it for context compaction. They cut latency from a hundred fifty seconds down to twenty-seven. That is such a huge developer-experience win.
Echo It really is. And it ties back to that thing we keep talking about… the control layer being the product. The model is just the engine, but the actual value here is the harness that lets you run dozens of calls per search request without the user noticing a lag.
Onyx Exactly. It's that same pattern. The loop is the product. They're even previewing a Mercury Router to handle the routing between open and closed models. It's just… it's all about the orchestration now.
Echo I mean, I'm still a little skeptical about the 'largest diffusion model ever' marketing, but the price point is hard to argue with. Twenty cents per million input tokens? Even without the launch discount, that's aggressive.
Onyx It's basically a race to the bottom on price, but a race to the top on speed. I love it.
Echo It's a weird spot to be in. But yeah, for a sub-agent that just needs to summarize a state or route a request, you don't need a giant frontier model. You need something that doesn't make the user wait.
Onyx Exactly. If you want to play with it, it's on the Inception API and Open Router. They're giving away a hundred million free tokens to get people started.
Echo I might actually try that. Just to see if the reasoning holds up under pressure.
Onyx Do it. I'll bet you it's faster than you think. Anyway, I'm starving… I'm gonna go find some food. Catch you later, Echo.