I Tested Mercury 3 (1068 T/S) — The Insanely Fast Agentic AI Model
Publicado no IA Dots ·
Agentes de IAAI SignalAssistente pessoaldot-ai_agentsdot-artificial_intelligenceidioma-enInglêsyt-a1T7sBJp0zU
Vídeo em Inglês, publicado pelo canal Codedigipt.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
The video introduces the Mercury 3 AI model from Inception AI, highlighting its high speed of 1,068 tokens per second, 1 million token context window, and performance compared to models like GPT-6 Luna and Claude Haiku 5.5. The presenter demonstrates using the model on its official platform and inside Cloud Code, covering use cases such as coding assistance, customer support automation, workflow automation, and real-time conversational AI.
Key points
- Mercury 3 achieves an output speed of 1,068 tokens per second on Nvidia GPUs.
- The model features a 1 million token context window, an upgrade from the 260k context in Mercury 2.5.
- Capabilities include reasoning, agentic coding, and long-context processing that match top frontier closed-source models.
- The model is available for free use on chat.inceptionlab.ai with an offering of 100 million free tokens.
- Practical use cases demonstrated include long-form text generation, logical reasoning challenges, code debugging, and code base exploration using Cloud Code.
Topics
Mercury 3 AI modelInception AIToken generation speedContext windowAgentic tasksCloud Code integrationAI coding assistant
Video transcript
Transcript from available captions; it may contain errors.
Hi guys, welcome back to another new exciting video, another new great interesting model, which is the Mercury 3. And it is from the Inception AI. I have tried this model inside the Cloud Code, and this model is pretty fast. You see, they have also mentioned on their official website, this model speed is 1068 token per second, means it is a instant level of speed. And also this model have beaten the GPT-6 Luna and Cloud Haiku 5.5. We know that Cloud Haiku 5.5 got released some days ago, and this model was actually a great model for the agentic kind of task, and that model speed was around 240 token per second. And GPT-6 Luna medium, its speed was 130 token per second. Basically, this GPT-6 Luna and Cloud Haiku 5.5, this model were was made for the high throughput and low latency agentic task. And now we have this Mercury 3, which is giving this 1068 token per second the output speed. And this model is actually really a great choice for the agentic task where we need the low latency and high throughput and high efficiency. And you see the pricing is also less than the others. And on this capabilities like this GPQA diamond, the reasoning capability, and also this agentic coding, and also the long context, on all of this, this model is giving the almost same kind of capabilities like the top frontier closed-source GPT-6 and Cloud. Now, the thing is this model is currently available for free to use inside their chat.inceptionlab.ai. This is the official platform. And also, currently they are giving the 100 million free tokens. Here you see that you can use. And I have tried it inside the Cloud Code, and I am actually happy with this model because the model is pretty fast, and also the model have great reasoning capability. And I have done the setup with the help of jan.ai. Previously also in my videos I discussed that how you can use that gen dot. Just go to our channel and here search gen and press enter and you will find this gen plus code cloud code setup, okay? If you watch this video, then you will get the idea that how you can connect any model with your gen desktop app. It is completely free, guys, and I am not promoting any kind of application. I just found it helpful, useful. That's why I am just giving you the suggestion, okay? And yes, I have mapped this uh Mercury 3 with the Sonnet. Here you see, this is the Sonnet model that I have mapped this Mercury 3. And uh how you can get the base model? Here you see, this is the base model. And for the KUI, just do the sign up on their platform. And after that, on the left-hand side, you will find this KUI section and just take that, use it. And uh why uh this Mercury 3 is very special? First of all, instant speed. On their official website, if you uh just uh scroll up below, here you see, 1,060 token per second on Nvidia GPUs. It is a great achievement. Next one is the context window. 1 million tokens context window. Previously, they had the 260k for the Mercury 2.5. Now, for Mercury 3, it is 1 million token context window. And quality. So, it is on par or better than the frontier speed uh optimized model like GPT-6, Luna, and Claude Haiku 5.5. So, now we have the alternative to the GPT-6, Luna, and Haiku 5.5. And also, the pricing they have kept uh currently, it is the 50% off as a promotional purpose. And also, agent friendly. It's low latency makes it useful for browser agent, real-time search, interactive tutoring, and customer support applications. Now, I have couple of prompts that I will try, but before that, let me just show you some of the use cases. Means, where you can try this. First of all, AI coding assistant. You can generate the code completions, boilerplate code, and unit test, SQL queries, and code explanation quickly. I am not giving you that suggestion that you can generate kind of complete production ready kind of code from scratch, okay? But, you can use this as a coding assistant. Means, you can actually generate some kind of unit test, SQL query, code explanation, okay? But, for high level of task, you have the model like Claude, Opus, and also GPT-6, Sol, right? But, this model is for the assistant purpose. Means, high volume text generation. Next is the customer support automation. Like, summarize the customer histories, drop replies, classify the incoming tickets. And wide fits potentially lower cost at scale. Next one is the AI agent and workflow automation. Like, you can generate the action plan, transform data, prepare the tool arguments, and summarize the results between steps in a workflow. Because of that low latency, it actually compounds across agent steps, okay? And next one is the real-time conversational AI. Like, power the voice assistant, interactive tutors, and conversational interfaces where the users expect quick responses. And also large-scale document processing, as it supports the 1 million context window. So, you can extract the structure information, categorize the document, summarize the reports, and produce the standardized output. Because it has the high throughput and the token also supports 1 million context window and low latency. So, that's why large-scale document processing can be very high efficient with this model, okay? And now, I have a couple of prompts that I want to test. Here you see, speed test you can do, like generate the long answer, okay? So, for generating the long answer, you can actually take this kind of prompt, like write a thousand word beginner-friendly guide, how to how AI agents work. So, I will just copy this and go to their official platform. First, I will try inside their official platform, then I will try inside the cloud code, okay? So, here you see I have selected the high version and uh this is the multi-utility three and I just paste it and press enter. Now, you will see Here you see that partial kind of text first is first it uh generated and after that it actually refined that and then we got the uh main text. So, if you just play this video slowly this portion when it was generating the test, then you you will notice that uh first it will actually uh generate some kind of partial text and then after that iteratively it is uh it is fine-tuning parallelly to give the final result. And you see within a second we got the answer, right? It's a kind of very very fast model. And the answer is a kind of great answer that we have got. Okay. Now, now next we will go for the next prompt. Let's say logical reasoning challenge, okay? Now, if I copy this and uh if I just paste it here and press enter. Now, you will see that first it will generate a deformed uh kind of text and uh partially uh formed representation. See, these are the partial formed representation, you see? Now, here it is thinking because it needs the logical uh reasoning. And here you see we got the answer. Within a second we got the answer. And here you see 6 second, okay? After 6 second we got the answer. Now, it has actually passed that real-world coding test. I will not do here. Uh debugging challenge, yeah, we can do that debugging challenge. Uh find every bug in this JavaScript code. As I told you that for code assistant purpose you can use it. Now, here you see it has opened the artifact also on the left-hand side. We have just uh found that what are the problems that is present inside this code and also what are the fix required, it has done that. You see a lot of explanation it has generated, but within a second. Right? Almost uh I think uh 2 to 3 seconds it took to generate all of this. So, this model is really very fast model, guys. And uh now what I will do, I will just uh go inside the cloud code. And uh here folder is uh this one. So, inside this folder I have actually uh couple of files. Basically, yesterday I tried the rocket launch uh this MiniMax M3.1 flash. This project I built. If you have not watched that that video till now, please watch that. This MiniMax M3.1 flash, that was the space bunny alpha model. Okay. And uh what I will do currently uh inside my folder I have these files. So, inside the JS folder we have this audio.js. Let's say I want to know what this audio.js contains. So, I will just ask it uh explain me the audio.js file. Press enter. So, now here you will see that it has the great identity capability. It is reading that audio.js. And uh I I will get the answer. Oh my god. How do you see? Almost it worked for the 12 seconds, and I got the answer. Right? Means what this audio.js is doing, it is a web audio API soundtrack or synthesizer no audio file. And lives on our window.launch.audio. And what it is doing, how do you see? What are the features that is present, and what are the controls that are present, all are in front of me. Means within a second you can actually explore any kind of code base. Now, if I check my uh balance currently, the number of tokens. So, first it was saying that zero, and I have used it. Now, let me refresh it. Okay, now it is saying that I have used, I think, 1 2 3 3 lakh uh 91,155 tokens. Means it is around 0.39% because they have given the 100 million free token, which is uh currently uh great number of tokens or you can say large number of tokens that they are giving. And you can actually explore with this uh token, right? And as the pricing is very low very less, so for this kind of agentic task, you can use it. And uh yes, guys, this is the information. And uh another thing, uh I actually generated one project uh idea, like what kind of uh task or what kind of project or where you can uh use this kind of model, this fast model, like AI operations agent for small business. So, let's say you are building a human supervised email and task automation agent that reads the incoming request, gather the relevant information, and update the business system, and prepare an appropriate response. Okay. And as this model is specifically version for the latency-sensitive agent and real-time applications, so uh this model can be a best uh choice for this kind of project. Okay. And here do you see, these are the project uh functionality, what the project does. It reads the incoming email and support the request, extract the customer intent, deadline, and important details, retrieve the company uh policies, and it asks for the human approval before sending the mails, record the result, and provide an audit trail. Like if the customer emails that my order has not arrived yet, can you check the status and let me know? So, what this agent will do, it will actually uh read the customer order number, find the status inside the system, and after that it will inform the admin that uh please update the status. Okay, so in this kind of scenario, you can use it. This model can be a great choice. As this model is very fast, and also the price very less than the top frontier model. Main important part is that this speed. Again, I will mention this speed. 1,068 is the initial level up speed that they are giving, okay? So, yes, guys. This is a thing that I wanted to discuss with you. And I hope you have found this video helpful detail explanation all of the steps that I have given you. I hope it will be very helpful for you to understand about this model. And if you have any question, just let me know in the comment section. And if you are visiting this channel for the first time, don't forget to subscribe this channel to get the this type of latest AI content daily. And see you guys in the next video. Thanks for watching. Bye-bye. Take care.
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, Assistente pessoal. Publicado no YouTube em 2026-10-20.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.
About Mercury 3
Mercury 3 is the newest model from Inception, a company that builds large language models powered by diffusion rather than traditional autoregressive generation. According to Inception, diffusion lets its models produce many tokens in parallel, which makes them several times faster and less than half the cost of conventional LLMs, and gives fine-grained control over outputs, such as following a specific schema.
Inception says its Mercury diffusion LLMs reach under 300 ms time to first token, 5 to 7 times higher throughput and up to 70% lower cost per task than speed-optimized frontier models. The company states that its team includes researchers and engineers from Stanford, UCLA, Cornell, Google DeepMind, Meta AI, Microsoft AI and OpenAI.
Fact sheet
- Model: Mercury 3
- Company: Inception
- Model type: Diffusion large language model (dLLM)
- Time to first token (Inception): Under 300 ms
- Throughput (Inception): 5–7× higher than speed-optimized frontier models
- Official site: Inception – official site
Read also
- Stop Building Apps. Build Tools for AI Agents
- Harness² : Google Open Sources for Self Improving AI Agents
- Mac Mini 16GB vs 32GB for AI Agents: OpenClaw Uses 1GB
Sources and references: Inception – official site · YouTube description of the video by the Codedigipt channel.