Inception Labs releases Mercury 2.5, a diffusion LLM serving 1,107 tokens per second

Inception Labs has released Mercury 2.5, the newest version of its diffusion-based large language model, which the company describes as its most capable production model to date. The announcement, which surfaced on the Hacker News front page, arrives after a period in which the previous Mercury 2 model was adopted by thousands of developers and put into production by dozens of enterprises, with usage growing more than tenfold according to the company.

The headline numbers

Inception Labs claims Mercury 2.5 is the largest diffusion language model ever trained, with a 40% intelligence increase over Mercury 2. The company places its quality alongside cost-optimized frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5 — a vendor-drawn comparison rather than an independent benchmark.

Other claimed specifications include throughput of 1,107 tokens per second on widely available NVIDIA GPUs, a 260K-token context window, and pricing of $0.20 per million input tokens and $0.75 per million output tokens. At launch the model is discounted by 80%, bringing those rates to $0.04 and $0.15 respectively. Feature support includes tunable reasoning, parallel tool calls and schema-aligned JSON output.

Shruti Koparkar, senior manager of product in NVIDIA's Accelerated Computing Group, credited the release with showing how quickly new architectures can mature into production-ready systems on NVIDIA's platform.

Shaped by production failures

Inception Labs says real-world workloads in search, voice and coding provided a stronger signal than benchmarks alone, and that customer feedback and production failure cases were used to sharpen evals and focus training for Mercury 2.5.

The announcement includes results from named customers. OpenCall, which builds AI phone agents, reports median model response latency close to 170 milliseconds on its production workload. Co-founder and CEO Oliver Silverstein said P99 response times fell from several minutes to about one second, and P50 dropped from 0.4 seconds to under 0.2 seconds, even including reasoning.

Augment Code uses Mercury for context compaction, model routing and MCP tool search. According to Inception Labs, moving compaction to Mercury cut latency by 82% — from roughly 150 seconds to 27 seconds — and reduced cost by 90% while maintaining quality, with tool-search summaries returning in under a second.

On the search side, the company notes that a single request in an agent or RAG pipeline can trigger dozens of model calls, from query planning and rewriting to reranking and summarization, and that several leading search-infrastructure companies now run Mercury in production, though it does not name them.

Previews: Mercury Voice and Mercury Router

Alongside the release, Inception Labs announced previews of two adjacent products. Mercury Voice is a diffusion LLM optimized for voice agents with the tightest latency budgets, targeting time-to-first-token under 170 milliseconds. Mercury Router analyzes incoming prompts with a diffusion model and routes them to open or closed models offering the best mix of quality, speed and cost.

Availability

Mercury models are available through the Inception API, Baseten and OpenRouter. Enterprise deployments add dedicated capacity, autoscaling, compliance controls and configurable data retention. New API users get 100 million free tokens to trial the model, and Y Combinator companies can claim $500,000 in deployment benefits. Inception Labs also says training has already begun on a larger successor, its biggest model yet, targeted for release in the coming months.

Why it matters

Diffusion language models generate text through iterative refinement rather than strictly left-to-right token generation, which is what enables throughput figures like 1,107 tokens per second. All quality and performance numbers above come from the vendor and its customers, so they await independent verification. Even so, Mercury 2.5 is evidence that the diffusion approach has moved past research demos into latency- and cost-sensitive production workloads — agent orchestration, voice interfaces, RAG pipelines — where per-call latency and cost compound across dozens of model invocations per user interaction. Its pricing also undercuts most comparable fast-serving models, adding pressure to an increasingly crowded segment of the market.