Inception, the startup building language models on diffusion rather than the usual token-by-token approach, has released Mercury 3. The company announced the model on Monday, October 19, 2026, calling it the most capable diffusion LLM released so far and the clearest evidence yet that the technique can scale.
Most large language models are autoregressive: they write one token, then the next, each depending on what came before. Diffusion models instead draft and refine many tokens in parallel, which suits GPUs well. Inception says Mercury 3 has more than 40 times the total parameters of Mercury 1, released in February 2025, and roughly three times its score across standard benchmarks. It has not disclosed the actual parameter count, which The Deep View noted makes direct size comparisons impossible.
The headline claims are about speed. Inception measures throughput at 1,068 tokens a second on NVIDIA hardware and says the model runs four to eight times faster than GPT-6 Luna, Claude Haiku 5.5 and GLM-5.3 Flash while matching or beating them on its chosen benchmarks, including agentic tests. The context window grows to 1 million tokens from 260,000 in Mercury 2.5, and the model supports tool calling, structured outputs and selectable reasoning effort. Speaking to The Deep View, chief executive Stefano Ermon put Mercury 3 in the same class as those Luna and Haiku models, only far quicker.
Pricing is set at $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount to $0.10 and $0.35 and cached input billed at a tenth of the normal rate. Inception has not said how long the discount will last; Ermon told The Deep View the company hopes to make the lower price permanent. The model is offered through an OpenAI-compatible API on Inception's own platform as well as OpenRouter, Baseten and Vercel, and OpenRouter's listing confirms the endpoint and pricing.
The company is aiming at latency-sensitive work: browser and search agents, voice and chat support, and tutoring. Early-access customers quoted by Inception include TinyFish AI, which says Mercury 3 completed half of 50 live web-navigation tasks in its own harness, and education startup RevisionDojo. The Deep View also reports that Inception used swarms of Mercury 3 agents to speed up the model by a further 16%, a claim the company has not documented in detail.
All of the performance numbers come from Inception or its customers. No independent leaderboard results were available at launch, the benchmark table is the company's own selection, and the weights are not open. The release lands as larger labs explore the same idea, with Google DeepMind having shipped DiffusionGemma in June, so independent testing will decide whether diffusion text models can keep pace on harder reasoning tasks.