Mercury 2.5: When an LLM Stops Writing One Token at a Time

Every major model you’ve used, GPT, Claude, Gemini, Llama, writes one token at a time, left to right, each one conditioned on everything before it. That’s why streaming responses trickle out word by word. Inception Labs just shipped Mercury 2.5, a diffusion language model that throws that constraint out, generating a whole block of text in parallel and iteratively refining it, and it’s now running at over 1,100 tokens per second on widely available Nvidia GPUs. That’s not a marginal speedup. It’s a different generation mechanism reaching a point where the intelligence gap with autoregressive models has closed enough to matter for real workloads.

Why generation order is the whole story#

Diffusion models don’t predict the next token, they denoise all of them at once. Start from a sequence that’s mostly noise or masked placeholders, then run a fixed number of refinement steps where the model looks at the entire sequence and improves its guess for every position simultaneously. Image diffusion models have worked this way for years; applying the same idea to discrete text tokens is harder because text doesn’t have the continuous pixel space that made denoising tractable for images, and Mercury is one of the first models to make that work at frontier scale. Because every step touches the whole sequence instead of one token, the number of forward passes stops scaling with output length, which is where the throughput jump comes from.

The generation reports 40% higher intelligence than Mercury 2, plus a 260K token context window. Inception’s own comparisons put that gain in the range of cost-optimized frontier models rather than treating diffusion as a speed-only tradeoff. Mercury 2.5 also adds tunable reasoning, parallel tool calls, and schema-aligned JSON output, features that read like a checklist for what production agent frameworks actually ask of a model, not a research demo flexing a new architecture.

Pricing is aggressive enough to invite testing. Standard rates are listed at $0.20 per million input tokens and $0.75 per million output tokens, with an 80% launch discount bringing that to $0.04 and $0.15 respectively. It’s available now through Inception’s own API, Baseten, and OpenRouter, so trying it doesn’t require picking a new primary vendor.

Try it: parallel denoising versus one-token-at-a-time#

Move the slider to run it.

Move the slider to run it.

Slide the steps down and the diffusion side still resolves the whole sentence, it just spends fewer refinement passes doing it, while the autoregressive side always needs exactly one pass per word no matter what you do. That’s the entire mechanical reason Mercury clears 1,100 tokens per second: the cost of generating a response stops being tied to how long the response is.

What I’d watch next#

The interesting question isn’t whether diffusion text models can be fast, Mercury already answered that, it’s whether the 40% intelligence jump holds up on harder reasoning benchmarks and agentic tool-use evals where autoregressive frontier models still lead. If it does, the real shift isn’t “diffusion versus autoregressive” as a permanent architecture choice, it’s that latency-sensitive products, voice agents, live coding assistants, anything where a user is staring at a cursor, get a serious reason to route traffic to a different generation mechanism entirely. I’d rather see that argued out on production latency dashboards than launch benchmarks, so that’s what I’m checking back on.