最新模型
Inception Releases Mercury 3, a Diffusion Language Model with One Million Tokens of Context
Inception launched Mercury 3 on October 19, expanding its context window to one million tokens and claiming output speeds of up to 1,068 tokens per second. Application developers have begun integrating it, but the company’s speed and quality comparisons still await independent evaluation.
Inception released Mercury 3, a diffusion language model, on October 19, increasing its context capacity from Mercury 2.5’s 260,000 tokens to one million. The company claims the new model can reach 1,068 output tokens per second on NVIDIA GPUs, and that its quality exceeds GPT-6 Luna and Claude Haiku 5.5 on some benchmarks. These comparisons should currently be treated as vendor test results. Official announcement
The key technical question is whether diffusion generation can maintain its speed as models scale. Inception says Mercury 3 has more than 40 times the total parameter count of Mercury 1. The company’s earlier technical explanation says diffusion generation uses multiple rounds of refinement to process multiple tokens in parallel, reducing the sequential dependencies of generating tokens one at a time. However, this principle does not by itself prove that inputs of up to one million tokens have similarly low latency. Mercury technical overview
Integration uses an OpenAI-compatible interface; the official example specifies mercury-3 through the chat completions endpoint. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with launch discounts cutting both prices in half. The official model page also says Mercury 2 and 2.5 are no longer available to new users. Eligible existing customers can access them through November 15, while enterprise contracts are handled according to their terms. Existing deployments will need to plan a migration. Model and migration information
There has been limited validation from an application developer. Curiosity announced its integration the same day and reported that its development workspace took about 0.4 to 1 second to stream the first text, while chat turns involving one or two tool calls took about 1 to 2 seconds. The team stressed that these were measurements from a single day in a single container, not a general benchmark. Curiosity integration notes
For engineering teams, the question worth testing is whether this can reduce total wait time for multi-step agents. Evaluations should measure time to first token, output speed, and tool round trips separately, and test quality with real Chinese prompts, long documents, and tool parameters. A figure of a thousand tokens per second is not enough to infer end-to-end application latency or success rates.