Speed is the feature people notice first in an AI app, and diffusion LLMs are the most interesting bet on speed right now. Inception released Mercury 3 on Oct 19, calling it "the most capable diffusion-based LLM released to date". Diffusion models came up about 1,500 times in our X sample this week.
Diffusion vs. "normal" LLMs, in one minute
Autoregressive models (GPT, Claude, Gemini and most others) write one token at a time. Each token waits for the one before it, so long answers take time.
Diffusion LLMs work more like image generators. They start from a rough draft of many tokens at once and refine it over several passes, in parallel. Fewer sequential steps means much higher throughput on the same hardware.
The open question has been whether diffusion models can get smart enough to matter. Inception's claim with Mercury 3 is that they can: in its words, "text diffusion scales".
Mercury 3 at a glance
All figures are from Inception's announcement:
| Mercury 3 | |
|---|---|
| Speed | 1,068 tokens/sec on NVIDIA GPUs |
| Context | 1M tokens (up from 260K in Mercury 2.5) |
| Size | 40x+ the total parameters of Mercury 1 |
| Price (list) | $0.20 / 1M input, $0.70 / 1M output |
| Price (launch, 50% off) | $0.10 / 1M input, $0.35 / 1M output |
| Cached input | 90% off the input price |
| Where | Inception API, OpenRouter, Baseten, Vercel |
| API style | OpenAI-compatible |
Inception says Mercury 3 is on par with or better than GPT-6 Luna and Claude Haiku 5.5 across benchmarks, while running 4–8x faster. Those are the vendor's own numbers. Speed-tier models are close enough that you should test on your own prompts. (New to Luna? Our GPT-6 explainer covers the lineup.)
Inception also offers 100 million free tokens to try the API.
Where speed actually matters
Inception's early-access partners describe use cases where waiting hurts:
- Browser and web agents. Agents call the model at every step, so delays compound. One partner, TinyFish, says Mercury 3 finished 50% of 50 live Mind2Web navigation tasks in its harness.
- Tutoring. RevisionDojo says hints and marked answers now come back "in about a second".
- Search on smart glasses. Octen reports a median generation time of 0.9 seconds and 2.2 seconds end to end.
- Voice and chat support. Inception also ships Mercury Voice, tuned for time to first answer token.
These are customer quotes published by Inception, not independent tests.
When to try it
Good fit:
- Agent loops with many short calls, where latency adds up
- Voice and real-time chat
- High-volume, cost-sensitive work, especially with cached prompts
Probably not your first pick:
- Your hardest reasoning or coding tasks. Inception compares Mercury 3 with speed-tier models, not top frontier models.
For pricing against other cheap, fast models, see our LLM API pricing comparison and MiniMax M3.1 Flash. For cutting costs on repeated prompts, read prompt caching explained.
Trying it
Because the API is OpenAI-compatible, you can usually point the OpenAI SDK at Inception's base URL with your Inception key. Get the exact base URL and model ID from Inception's API docs, as we haven't reproduced them here. Mercury 3 is also on OpenRouter if you already use it to route between models.
Benchmark it on the task you actually care about. Measure latency and output quality, because a fast wrong answer is still wrong.
