Not just a model - It's a diffusion model. Instead of Next Token prediction it builds the entire page at once and then denoises it. Kind of mind blowing when you watch it happen.
"If you do not want us to use your User Submissions to train our models, you can opt-out by setting the ‘Improve the model for everyone’ option under User Settings in the API Platform to OFF."
Inception is one of the most interesting neolabs with their diffusion-based architectures. My understanding is that their primary business is low latency voice applications but they are seriously pursuing coding.
We tested Mercury 2.5 Preview, which is nowhere close to the frontier (and not advertised as such), but it's actually usable as a general-purpose chatbot. It's comparable in problem solving ability to some last-gen open weights models, and the price and cost make it compelling. However, they have not figured out general purpose tool use and agentic coding (their model performs worse on our problems when given a custom harness). If they do, I see a lot of real-time applications that the speed and cost will enable.
Same. Models that are fast (and thus by extension, in some direction, efficient) but not local/self-hosted IMO fill a niche for quick and cheap (and acceptable quality, of course) inference in business contexts.
This should make an excellent choice for arbiter in llm-consortium, mercury-2 was pretty good. One of the main drawbacks of the multi-model system is the added latency of the llm judge, but having a model run at 1100tps goes a long a way to alleviate that.
I used it for testing my voice agent. It was basically what I expected. Good fast model but "generic" or "vanilla" is how i would describe its personality emulation capability as. Gemma models still outperform it in that department. As far as technicals, one thing i found annoying is cash use was not that good, it missed more then i liked, i contacted support and they were fast and responsive and said they were working on that issue, maybe they solved it with 2.5? Anyways, im prolly gonna try 2.5 again see if anything different, but cant deny the speed, thats the biggest thing this company has going for this offering as if you are in the business of classical cascaded voice agent systems, latency is number one priority and this thing is fast....
latency, instruction adherence, reliable tool cools, conversationality are all in tension.
it's great when you can get a 170ms ttft. but if you have 700 ms endpointing on the stt side and 300ms ttfb on the voice side, then you haven't really made something super snappy.
I'd imagine at this point they are likely an acquisition target if they can get a halfway decent model. I can't imagine having diffusion sub-agents (or sub-sub-agents) in an orchestration wouldn't be beneficial.
I'm less bullish, this model and previous are halfway decent and you can do diffusion sub-agents and sub-agent-agents today, and there hasn't been a sea change, or anything noticeable or well-known. Economically, there's ~no moat, diffusion models aren't a mysterious untame-able force.
interesting:
Quality: 40% increase in intelligence from Mercury 2. Comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.
available : https://openrouter.ai/inception/mercury-2.5
They are comparing it to 2 and 3 version old flash/fast versions of models but purely for tok/s. Then only comparing it to Mercury 2 on intelligence. This is very misleading and I suspect this model is basically useless.
Oh great a new model annoncemzzzzzz ZZZZZZZZZ
Not just a model - It's a diffusion model. Instead of Next Token prediction it builds the entire page at once and then denoises it. Kind of mind blowing when you watch it happen.
Got my hopes up when it said widely available GPUs that it would be open weights but it doesn’t seem like it sadly
Not related to the Mercury language:
https://mercurylang.org/
I like the model.
FYI:
"If you do not want us to use your User Submissions to train our models, you can opt-out by setting the ‘Improve the model for everyone’ option under User Settings in the API Platform to OFF."
This token speed is too fast.
Inception is one of the most interesting neolabs with their diffusion-based architectures. My understanding is that their primary business is low latency voice applications but they are seriously pursuing coding.
We tested Mercury 2.5 Preview, which is nowhere close to the frontier (and not advertised as such), but it's actually usable as a general-purpose chatbot. It's comparable in problem solving ability to some last-gen open weights models, and the price and cost make it compelling. However, they have not figured out general purpose tool use and agentic coding (their model performs worse on our problems when given a custom harness). If they do, I see a lot of real-time applications that the speed and cost will enable.
the output is good and fast. like it, not only fast, but a new architectures.
Congrats! Happy to see someone seriously pursuing this direction.
Same. Models that are fast (and thus by extension, in some direction, efficient) but not local/self-hosted IMO fill a niche for quick and cheap (and acceptable quality, of course) inference in business contexts.
This should make an excellent choice for arbiter in llm-consortium, mercury-2 was pretty good. One of the main drawbacks of the multi-model system is the added latency of the llm judge, but having a model run at 1100tps goes a long a way to alleviate that.
Anyone here use Mercury 2.0? Curious what your experience with the model is.
I used it for testing my voice agent. It was basically what I expected. Good fast model but "generic" or "vanilla" is how i would describe its personality emulation capability as. Gemma models still outperform it in that department. As far as technicals, one thing i found annoying is cash use was not that good, it missed more then i liked, i contacted support and they were fast and responsive and said they were working on that issue, maybe they solved it with 2.5? Anyways, im prolly gonna try 2.5 again see if anything different, but cant deny the speed, thats the biggest thing this company has going for this offering as if you are in the business of classical cascaded voice agent systems, latency is number one priority and this thing is fast....
latency, instruction adherence, reliable tool cools, conversationality are all in tension.
it's great when you can get a 170ms ttft. but if you have 700 ms endpointing on the stt side and 300ms ttfb on the voice side, then you haven't really made something super snappy.
I'd imagine at this point they are likely an acquisition target if they can get a halfway decent model. I can't imagine having diffusion sub-agents (or sub-sub-agents) in an orchestration wouldn't be beneficial.
I'm less bullish, this model and previous are halfway decent and you can do diffusion sub-agents and sub-agent-agents today, and there hasn't been a sea change, or anything noticeable or well-known. Economically, there's ~no moat, diffusion models aren't a mysterious untame-able force.
interesting: Quality: 40% increase in intelligence from Mercury 2. Comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. available : https://openrouter.ai/inception/mercury-2.5
I like the text output on logical and historical content that I sampled so far
They are comparing it to 2 and 3 version old flash/fast versions of models but purely for tok/s. Then only comparing it to Mercury 2 on intelligence. This is very misleading and I suspect this model is basically useless.
I love this it is so fucking fast!
I thought this was going to be about a boat motor.
I thought it was going to be about Mercury[1] the programming language, and was confused about the domain name.
[1] - https://www.mercurylang.org/