No more Claude or ChatGPT - Local AI on your GPU + DaVinci Resolve Studio 21.1 - MCP Server

Vídeo em Inglês, publicado pelo canal Team 2 Films.

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

This video explains how to run large language models locally and connect them to DaVinci Resolve using an AI agent, detailing hardware requirements, setup processes, and practical editing workflows.

Key points

• Reasons for running local LLMs include economy, ethics, legal compliance, and speed.
• Running LLMs requires sufficient VRAM for model weights and context, utilizing techniques like quantization and offloading.
• A dual-GPU setup allows DaVinci Resolve and the local LLM to run on the same computer without resource conflicts.
• The demonstration uses Strata as an inference engine, the Qwen parameter model, Hermes as the AI agent, and DaVinci Resolve's native MCP integration.
• Local AI agents can generate O-graph animations based on user prompts and import them directly into DaVinci Resolve.

Topics

Local LLMDaVinci ResolveAI AgentsGPU and VRAMO-graph AnimationsHardware SetupQuantization

Video transcript

Transcript from available captions; it may contain errors.

In our previous video, we demonstrated DaVinci Resolve 21.1's native MCP integration with AI agents using Claude. >> [music] >> In this video, we will explore how to take advantage of locally run large language models rather than using cloud-based services [music] like Claude or ChatGPT. There are several reasons why someone might choose to run their own LLM locally. Number one, economy, but of course only once you've passed the critical payback point. Number two, ethics. Some users would prefer not to support the development of large data centers or AI behemoths. Number three, legal. When working on certain projects, some NDAs specifically prohibit the use of cloud-based services. And number four, speed. Dedicated hardware might have some speed advantages. This is not an AI masterclass, but we will cover some basics. LLMs are run on your computer's GPU, not its CPU, because a GPU's architecture runs thousands of threads simultaneously, whereas a CPU might only manage 32 or fewer. The critical spec for running LLMs is VRAM. A faster GPU will increase the speed at which your LLM will run, but if you don't have enough VRAM to load the model, you can't run it at all. At the core of any LLM are weights. They are billions of numbers that the LLM uses to process inputs and predict outputs. Those weights must be loaded into your VRAM. Attempting to access them from your SSD or even your computer's main RAM would add crippling latency. The size of LLMs are typically measured in billions of parameters. A 27 billion parameter model with weights stored with 16-bit precision would require about 54 GB of VRAM. Reducing the precision of those weights to 8-bit barely impacts the quality of the model, but it halves the required VRAM to 27 GB. 4-bit precision halves the required VRAM again, but starts to have an impact on the quality of the model. A higher number of parameters and a greater amount of precision typically increase the efficacy of a model. In addition to storing the weights, you also need room in your VRAM to store context. If the weights are your model's long-term memory, then context is its short-term memory. The more context you allow your model, the greater the complexity of the tasks it can perform. It's possible to combine multiple GPUs to increase the amount of VRAM you have. Some models are designed to use a process called offloading where only the parts of your model that are currently being used are kept in VRAM while other parts are parked on your system RAM and loaded as needed. That means affordable GPUs can be used to run LLMs. We'd recommend a minimum of 16GB of VRAM. If you use a model that utilizes offloading, having lots of system RAM is also needed. Apple's unified memory architecture means the GPU can borrow up to 75% give or take of system memory, which makes Macs very versatile for running LLMs. In this demonstration, I'm going to be using my ASUS ProArt Windows system. It has an NVIDIA 5080 with 16GB of VRAM. I also have a 5060 Ti on hand that also has 16GB of VRAM. Well, these two cards are capable of loading the same models. It's just that one will work quicker than the other. It's good to know that the same models can be run at different price points. Running DaVinci Resolve and an LLM on two separate computers has many advantages, but I'm going to show you how you can run DaVinci Resolve and an LLM on the same computer. Here's our first problem. Both DaVinci Resolve and an LLM make heavy and exclusive use of the GPU and its VRAM. To circumvent that, you can install more than one card into your system. Ideally, I would run two 5080s, but GPUs are expensive, so use what you have. This 5070 is half the cost of a 5080 and it has 12GB of VRAM. I'll use the 5080 for the LLM and the 5070 for Resolve. I'm able to run two graphics cards because my motherboard, the ProArt X870E Creator, has two PCIe 5.0 * 16 slots that run at * 8 with both cards installed. And the case in which I'm operating has sufficient power and cooling to sustain two cards. Whether you are operating on one computer or have additional hardware that will allow you to host your LLM on a second computer, you will still need two graphics cards. You need to choose an LLM that fits within your VRAM, but also leaves sufficient room for context. It's beyond the scope of this video to summarize all the different LLMs available for download. We're using Strata, an open-source inference engine that uses the Qwen 3.8 FlashNext 125 billion parameter model. It [snorts] makes use of offloading and quantization to reduce the size of the model. It will happily operate on a 16 GB graphics card, but it requires a large amount of system RAM for offloading the model. I'm using 128 K of context. My system has 96 GB of RAM. I expect Qwen to use around 60 GB of system RAM and to use all of the 16 GB of VRAM. Setup is incredibly easy. Just follow the prompts in the start here script. An AI agent is needed as a go-between to interact with the LLM and DaVinci Resolve. The agent has tools that allow it to interact with your computer. I'm using Hermeys. Giving access to Qwen is as simple as creating a custom endpoint and providing it with Qwen's local address. This video is sponsored by Audio, where with our discount code, you can get 70% off an entire year of royalty-free music. We use Audio for all of our music. I'd like to highlight a genuinely useful feature, and that is their phone app. The homepage features new artists and playlists. That's a great place to start if you don't know what you're looking for yet. If you're out and about when you've got time to kill, instead of listening to the music you already know, you can explore Audio's library and start building ideas for your next edit. If you're looking for something specific, head to search. You can do all the normal stuff like search through song titles, artists, and keywords, or you can hit this button in the middle. This is Hermes AI. It [music] understands the music in Audio's library. You can either describe the kind of song you're looking for, or describe the kind of video you're looking to make. Either way, when you find songs you like, mark them as a favorite. Now when you sit down to start editing, instead of starting from zero, you're already ahead. You can get an entire year of Audio Pro for just $60 [music] using our discount code in the description. In Resolve's system preferences, I have configured it to use the 5070. I'll ask Hermes if it can access Resolve's MCP. Connect to DaVinci Resolve's native MCP server. We now have a completely local setup. DaVinci Resolve doesn't use any cloud-based AI, and now we have it connected to a local AI agent that also doesn't use any cloud compute. To demonstrate this installation, I will ask Hermes to create a dot O graph animation for DaVinci Resolve. Resolve already has the amazing Fusion page for creating graphics and animations, but in DaVinci Resolve 21, support for O graph animations was added. O graph is an open standard for broadcast graphics created by the European Broadcasting Union. It uses HTML, CSS, and JavaScript, and other web-based technologies to build graphics for videos. Why might this be of interest? Because LLMs have got really good at writing HTML. So let's ask Hermes to create me one. Create a lower third name and title O graph with animations for DaVinci Resolve and import it into my DaVinci Resolve media pool, and then insert it into a timeline. Check it works. It should have user editable fields. If you're unsure how to create O graphs, here is a folder of other graphics that you can learn from. And I'm pointing it to a folder on my desktop that has O-graphs in it. This is the AI agent's first time creating an O-graph for me. So, it will take a little bit longer. Once Hermes has worked this out, it will create a skill that outlines basic steps that it can recall in the future and do this more quickly. And here is my editable O-graph in DaVinci Resolve. I have controls for changing the text and colors. That took 6 minutes to produce. And remember, it's longer the first time. Now, let's throw it a harder challenge. This prompt will be quite long. I'm going to ask it to create an animation that shows how the user, Kwen, Hermes, and DaVinci Resolve interact with each other. That's quite a detailed prompt. This isn't the first time I've done this, so I've learned a few lessons along the way regarding what I need to be specific about in my prompts. The reality is that for you to create this graphic from scratch would likely involve several rounds of prompting as you adjust the graphic each time. This process does not replace people. It's a tool for people to use. Hermes has also learned lessons from creating the first O-graph. Regardless, this will take a while. Stratus's interface shows my system utilization. The GPU is at 100% because of offloading, Kwen is actually using around 50% of my CPU and 60 GB of system RAM. But there's more than enough CPU performance and more than enough system RAM available to run DaVinci Resolve. And thanks to that second GPU, I shouldn't experience any major performance issues. To confirm that, here's a 12K B-roll clip playing with three frames of temporal noise reduction applied in a UHD timeline, and I'm still getting 24 frames per second playback. This animation took 16 minutes to create, and it has all the features I requested. That 16 minutes also includes time spent screenshotting the animation and validating that it met my requirements. Thanks to the second GPU, I was able to work in the meantime. Here's some thoughts. The only software that I've paid for is DaVinci Resolve Studio. Now, this does require moderate hardware. Services like Claude and ChatGPT are likely more capable, [music] but it's not a night and day difference. It's really impressive what is possible with a free, locally run large language model. Here's some pros. [music] Creating a simple AI workflow graphic manually would have taken me perhaps an hour. Now that it's quicker, I might include more graphics. And this process will only get better as more capable LLMs are released. It's very easy for me to switch out [music] Qwen for a newer or more capable LLM. Here's some cons. [music] I'm losing some intentionality as decisions are being made for me. Uh granted, the more precise I am with my prompts, the more closely the animation resembles what I imagined, but that does require me to have the skill of being able to describe my ideas to someone else, or in this instance, [music] the AI agent. It's important to note, I'd still sacrifice some intention if I were to build this myself manually because often the aesthetic or function of an animation is heavily influenced by the tools that I'm familiar with, and the ones that are most efficient to use. [music] Also, because I don't understand the way this graphic has been made, it's harder for me to edit it. I'm mostly stuck using prompts [music] if I want it changed. Although I could ask the agent to make me a tool for editing O-graphs, but now I'm spending time developing software rather than telling stories. I think [music] that I would make a more creative and aesthetically pleasing animation if I were to do it myself. But to create an animation takes time. So, it has to be important for me to make one. A good plus of that is that there are fewer animations in my videos, which gives the ones that are in there more impact. Here's my takeaway. [music] If I start to use these tools in my work, I want to make sure I don't sacrifice any of my intentionality, and that I don't overuse these tools. Story always comes first and at the end of the day, I'm a human being using videos to communicate with other human beings. Technology facilitates that process, [music] but it shouldn't drive that process. If you're interested in creating your own local LLM setup, check out ASUS's range of graphics cards. Don't forget you need a studio license of DaVinci Resolve. There's a plethora of other GPUs, LLMs, and agents that you could build your own local setup with. Not to mention the advantages of hosting your LLM on a different computer. If you're looking to purchase computer equipment, get in touch with our friends at Scan. If you're interested in the equipment that we use to film our videos, go visit our friends [music] at Pro AV. As always, we'd love to hear your balanced thoughts in the comments. If you enjoyed [music] this video, please give it a like and consider sharing it with a friend. And if you didn't enjoy this video, don't worry, there are more videos about film, cameras, and production coming soon. Make sure you're subscribed so you don't miss out on future videos. Thanks for watching.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal. Publicado no YouTube em 2026-10-16.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.