Inception Labs announced on Monday, October 19, 2026, Mercury 3, described by the company as the most capable diffusion LLM released to date. Unlike traditional autoregressive models that generate tokens one by one, Mercury 3 produces multiple tokens in parallel, enabling speeds above 1,000 tokens per second on standard NVIDIA GPUs.
What happened?
According to the official Inception blog post and the Business Wire release, Mercury 3 is more than 40 times larger than Mercury 1 in parameters and about 3 times more intelligent on standard benchmarks, while maintaining the same order of speed. The context window reached 1 million tokens, up from 260,000 in Mercury 2.5.
The company states that the model matches or surpasses GPT-6 Luna and Claude Haiku 5.5 in quality across various benchmarks, including IFBench, GPQA Diamond, and SWE-Bench Atlas, while being 4 to 8 times faster. These results are reported by Inception itself and await broader independent evaluations.
At launch, pricing is 50% off: $0.10 per million input tokens and $0.35 per million output tokens (regular price $0.20 and $0.70). Cached tokens receive an additional discount. The model is available on the Inception API, OpenRouter, Baseten, and Vercel, with OpenAI API compatibility.
Why it matters
The main open question in the community was whether diffusion LLMs (dLLMs) could scale without losing their speed advantage. Mercury 3 suggests they can. Co-founder and CEO Stefano Ermon, a Stanford professor and co-inventor of techniques such as DDIM and score-based models, highlighted that diffusion opens a path to scale intelligence without a proportional increase in latency.
Early-access partners are already using the model in web navigation agents, real-time tutoring, customer support, and search. In internal web-agent tests on Mind2Web, Mercury 3 completed more tasks than GPT-6.1-Sol with similar token volume and number of steps, at a lower per-token cost.
Inception's portfolio also includes Mercury Voice (optimized for low-latency voice) and Mercury Decide (for high-volume classification). The launch strengthens an alternative to the dominant autoregressive models from OpenAI, Anthropic, and Google in latency-sensitive scenarios.
Sources
- Official Inception announcement: Diffusion at Scale: Introducing Mercury 3
- Business Wire release
- Stefano Ermon's thread on X
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for full context.
By GeekikiBot
