Mercury 3

A language model that does not write left to right

Almost every chat model you have used, from ChatGPT to Claude, generates text the same way: it predicts the next token, appends it, and predicts the next one. That is called autoregressive generation, and it has a built-in speed limit, because token 500 cannot be produced until token 499 exists.

Mercury takes the approach image generators use. A diffusion model starts from a rough, noisy draft and refines it over a series of steps. Applied to text, that means the model works on a whole block of tokens at once and sharpens all of them together, instead of committing to one word at a time. Fewer sequential steps per answer is what produces the speed. The tradeoff is that the technique is less mature for language than autoregressive generation, and it has been an open question whether it keeps up on harder reasoning as models get bigger.

Mercury 3 is Inception’s answer to that question. Released October 19, 2026, it is the company’s largest model, more than 40 times larger than Mercury 1 by Inception’s account (it does not publish an exact parameter count). Inception presents it as evidence that text diffusion scales.

What Inception claims

Inception reports 1,068 tokens per second on NVIDIA GPUs and says Mercury 3 runs 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5 while matching or beating them on 8 of the 12 benchmarks it reported. On a set of 50 live Mind2Web browser-navigation tasks run in Inception’s own web-agent harness, it says Mercury 3 completed half.

All of these are vendor-run. Inception chose the benchmarks, the comparison models, the hardware, and the harness, and its announcement does not publish per-benchmark scores for the headline comparison. That does not make the claims wrong, but it means they describe Mercury 3 under Inception’s conditions. “Over 1,000 tokens per second” is a measurement in a particular setup, not a guarantee you will see through every provider.

The comparison models also tell you where Mercury 3 sits. GPT-6 Luna and Claude Haiku 5.5 are the fast, inexpensive tiers of their families, not the flagships. Mercury 3 is positioned as a faster option in that tier, not as a rival to the strongest models.

Specs and access

Enterprise customers can get dedicated capacity, compliance controls, and configurable data retention through Inception’s sales team. The same company also offers Mercury Voice for low-latency voice apps and Mercury Decide for high-volume classification.

When speed actually matters

A faster model is not automatically a better choice. Speed pays off in specific places:

• Agent loops. An agent that makes 40 model calls to finish a task waits on each one. Cutting each call from several seconds to under one changes how long the whole job takes.
• Interactive tools. Autocomplete, inline editing, and live search answers feel different when the response arrives almost at once.
• Voice. Latency between a person finishing a sentence and the reply starting is the main thing people notice. See latency.
• High-volume pipelines. Extraction, tagging, and summarization jobs over large document sets finish sooner and tie up less capacity.

It matters less when a task needs one careful answer and you would happily wait 30 seconds for it to be right.

Cost is more than token price

Mercury 3’s list price is low, and the 90% cache discount helps for agents that resend the same context. But per-token price is only one input. A fast model that needs two retries to get something right can cost more, and take longer, than a slower model that gets it right first time. The number worth tracking is cost per completed task, which includes retries and failures. Mercury 3 is a natural candidate for the worker tier in a model routing setup, with harder steps escalated to a stronger model.

Limitations

• Vendor benchmarks only, so far. No independent full evaluation had been published at launch.
• Unclear on tool use and coding specifics. Inception’s launch post does not detail function-calling support or coding benchmarks, so test these if they matter to you.
• Not the frontier tier. It is compared against fast, cheap models, not flagships.
• No weights. You cannot self-host it.
• Newer technique. Diffusion text generation has a shorter track record, and failure modes may differ from what you are used to with autoregressive models.

Alternatives

• DiffusionGemma, Google’s experimental open-weights diffusion text model, if you want to run the technique yourself
• MiniMax M3.1 Flash, a low-cost open-weight model for agent workers
• DeepSeek V4.1 Flash, a cheap open-weight execution model
• Gemini Flash models and GPT-6 Luna, the fast tiers Inception compares against

For how to combine a fast model with stronger ones, see Why You Need More Than One AI Model and Planner-Worker Agents.