Inception has launched Mercury 2.5, a diffusion large language model (dLLM) that claims over 1,100 tokens per second in production, positioning it as the fastest reasoning LLM in operation. This release comes amidst a crowded September 2026 model launch window, with Inception betting on architecture and speed to differentiate.
TL;DR
- Inception's Mercury 2.5 dLLM achieves over 1,100 tokens per second in production, claiming the title of fastest reasoning LLM.
- The model's diffusion architecture generates text in parallel, potentially reducing AI infrastructure costs for high-throughput workloads.
- Mercury 2.5 launches alongside preview products Mercury Voice and Mercury Router, hinting at future voice interface and task-based model orchestration.
What happened
Inception introduced Mercury 2.5 on September 8, 2026, via a Business Wire announcement. The model is touted as the fastest reasoning LLM in production, with a throughput of over 1,100 tokens per second. This claim is based on Inception's diffusion LLM architecture, which generates text in parallel rather than sequentially.
Accompanying Mercury 2.5 are two preview products: Mercury Voice, a voice interface, and Mercury Router, a task-based model orchestrator. Inception has not yet published technical specifications, pricing, or availability dates for these preview products.
Why it matters
Mercury 2.5's high throughput could significantly reduce AI infrastructure costs for enterprises running high-volume, latency-sensitive AI features. This includes applications like live chat support, in-editor code completion, and voice assistants.
The model's diffusion architecture may not match autoregressive models in long-form coherence but excels in latency-sensitive, high-throughput workloads. This positions Mercury 2.5 as a cost-optimization layer for companies looking to control AI compute bills.
Inception's focus on architecture and speed differentiates it from other September 2026 model launches, such as the Institute of Foundation Models' open-weight models and iFlytek's large reasoning model. This diversification in the foundation model market highlights the increasing specialization among AI labs.
Key facts
- Mercury 2.5 achieves over 1,100 tokens per second in production, according to Inception.
- The model is based on a diffusion LLM architecture, which generates text in parallel.
- Mercury 2.5 launched on September 8, 2026, alongside preview products Mercury Voice and Mercury Router.
- Inception claims Mercury 2.5 is the fastest reasoning LLM in production.
- The model is positioned to compete with other September 2026 launches, including the Institute of Foundation Models' K2 Horizon and iFlytek's Spark X2.5.
- Mercury 2.5 is designed to reduce AI infrastructure costs for high-throughput workloads.
- Inception has not yet published pricing, context window size, or availability dates for Mercury 2.5 or the preview products.
Context
Diffusion models originated in image generation and have since been adapted for text. Inception has built its identity around commercially viable diffusion LLMs, with Mercury 2.5 being the latest evolution in this family.
The September 2026 model release window has seen a surge in AI model launches, with labs focusing on specific technical wedges such as openness, specialization, or architecture. This diversification reflects the maturing AI market and the increasing specialization among AI labs.
Independent, standardized throughput comparisons across Mercury 2.5, Spark X2.5, and GPT-6 Astra have not yet been published. Until third-party evaluators run their own tests, the 1,100 tokens-per-second figure remains a vendor-reported number.
