Inception Labs released Mercury 2.5 on Tuesday, a new version of its commercial diffusion language model that the company describes as the most capable diffusion LLM on the market and, to its knowledge, the largest diffusion language model ever trained. According to the company's announcement, the model delivers a 40% increase in intelligence over its predecessor, Mercury 2, while retaining the same low-latency, low-cost serving profile that has made the diffusion architecture attractive for high-volume workloads.
The company, led by CEO Stefano Ermon, claims Mercury 2.5 is comparable to cost-optimized frontier models including GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5. Those comparisons are vendor-reported and have not yet been verified by independent benchmarks, but they position a diffusion model — long a research curiosity — as a production alternative to the autoregressive incumbents. For the latest AI model news, follow AI Buzz Wire's ongoing model coverage. latest AI model news
Speed, Context and Price
The headline numbers are aggressive. Inception Labs says Mercury 2.5 generates 1,107 tokens per second on widely available NVIDIA GPUs, with a context window of 260,000 tokens. Pricing is set at $0.20 per million input tokens and $0.75 per million output tokens, with a launch promotion discounting the model 80% to $0.04 per million input and $0.15 per million output.
On the capability side, the model ships with tunable reasoning — letting developers trade latency for depth on a per-request basis — along with parallel tool calls and schema-aligned JSON output for structured generation. The company frames the release as the first result of a training loop driven by production telemetry: since Mercury 2 launched, thousands of developers have built with it, dozens of enterprises have deployed it, and usage has grown by more than an order of magnitude.
Why Diffusion Models Generate Text Faster
Mainstream LLMs from OpenAI, Google and Anthropic are autoregressive: they generate text one token at a time, with each token conditioned on everything generated before it. That sequential dependency puts a hard floor under generation speed. Diffusion language models instead start with a block of noise and iteratively refine all tokens in parallel, which allows far higher throughput on conventional GPU hardware once the model is trained to converge cleanly.
The trade-off has historically been quality: diffusion models have trailed autoregressive ones on reasoning and instruction-following benchmarks. Inception Labs' core bet — and the claim underlying Mercury 2.5 — is that this gap has narrowed to the point where diffusion wins on the axis most customers actually feel: latency and cost at production volume.
Production Claims From Early Customers
The announcement leans heavily on customer deployments rather than benchmark tables. OpenCall, which builds AI phone agents for live customer calls, reported that moving to Mercury brought its median model response latency to roughly 170 milliseconds. Oliver Silverstein, OpenCall's co-founder and CEO, said the company's P99 response time dropped from several minutes to one second, and its median from 0.4 seconds to under 0.2 — figures he described as significantly faster than any other provider the company had tested, including with reasoning enabled.
Augment Code, an AI coding assistant maker, uses Mercury for context compaction, model routing and MCP tool search. According to Inception Labs, moving compaction workloads to Mercury cut that step's latency by 82%, from roughly 150 seconds to 27 seconds, and reduced its cost by 90% while maintaining quality. The use case illustrates where the economics bite: coding agents make many small supporting model calls around each primary generation, and latency and cost compound across those calls.
NVIDIA, whose hardware runs the model, also weighed in. Shruti Koparkar, a senior manager of product in NVIDIA's Accelerated Computing Group, said Inception has advanced diffusion-based language models on NVIDIA AI infrastructure, and that Mercury 2.5's quality step-up with sustained speeds shows how quickly new architectures can mature into production-ready systems.
Mercury Voice and Mercury Router Previews
Alongside the model, Inception Labs announced previews of two new products. Mercury Voice is a diffusion LLM optimized for voice agents with the tightest latency budgets, advertising time-to-first-token under 170 milliseconds. Mercury Router uses a diffusion model to understand incoming prompts and route them to whichever model — open or closed — offers the best mix of quality, speed and cost, positioning Inception as a layer above its competitors as well as one of them.
Mercury 2.5 is available through Inception's own API as well as through Baseten and OpenRouter. Enterprise deployments add dedicated capacity, autoscaling, compliance controls and configurable data retention. The company is offering 100 million free tokens to developers trying the API.
The company also said it has already begun training its next model — its largest yet, targeted for release in the coming months — which it describes as a leap in capability that gives up none of diffusion's speed or token efficiency. If that claim holds, the pressure on autoregressive incumbents will be felt first in the commodity tiers of the market, where price and latency decide most contracts.
Stay Ahead of AI
Follow the latest AI developments and breaking artificial intelligence news as new model architectures race into production.
Read more AI news →
