CPU - Intel i9-14900K (24 cores, 32 threads)
GPU - RTX 4090 24GB + RTX 4070 12GB (K43SG eGPU dock)
RAM - 64GB DDR5
Already running - Ollama (38 models, up to 181 tok/s), ComfyUI, Whisper STT
$ I keep a list of things I can run locally. It started as a short list. It is no longer a short list.
Every capability below runs on one desktop. Not a data center. Not a cloud instance billed by the hour. A PC under a desk, drawing maybe 500W at full tilt, with parts you can buy at Micro Center. Some of these would have cost six figures in compute two years ago. Today they run between gaming sessions.
Here are 14 things this machine can do - ranked by how hard they made me stare at my monitor.
Every capability below runs on one desktop. Not a data center. Not a cloud instance billed by the hour. A PC under a desk, drawing maybe 500W at full tilt, with parts you can buy at Micro Center. Some of these would have cost six figures in compute two years ago. Today they run between gaming sessions.
Here are 14 things this machine can do - ranked by how hard they made me stare at my monitor.
Mind-Blowing
Things that genuinely should not be possible on hardware you can carry upstairs.
1
Fine-tune a 70B parameter model on your PC
4090
Unsloth with QLoRA fits a 70B model into 24GB VRAM by quantizing the base weights to 4-bit and training only low-rank adapter layers. An overnight run - 8-12 hours - produces a fine-tuned model you can export to GGUF and load straight into Ollama.
"I trained a 70 billion parameter model on my gaming PC."
That sentence was science fiction three years ago. Now it's a weekend project. You can give it your writing samples, your code style, your domain knowledge. The model learns. You export it. You run it locally forever. No subscription. No API. Just weights on your disk.
2
Real-time AI video style transfer
4090
StreamDiffusion and Daydream take a live webcam feed and re-render every frame through a diffusion model in real time. Your face becomes anime. Your room becomes an oil painting. Your dog becomes a watercolor.
At 15-30 FPS with sub-100ms latency, it's smooth enough to pipe through a virtual camera into OBS and onto a Zoom call. Your coworkers see you as a Studio Ghibli character. They do not know it's running on the box under your desk.
At 15-30 FPS with sub-100ms latency, it's smooth enough to pipe through a virtual camera into OBS and onto a Zoom call. Your coworkers see you as a Studio Ghibli character. They do not know it's running on the box under your desk.
3
Voice AI that crosses the uncanny valley
4090 4070
The latest generation of voice models doesn't just sound human - it sounds like a specific human having a conversation.
Sesame CSM-1B generates natural speech with "umms", "uhhs", and those micro-pauses that make you forget you're listening to a machine. Dia2 produces multi-speaker dialogue with distinct voices, emotion, and timing. CosyVoice2 does streaming synthesis at 150ms first-chunk latency - fast enough for real-time conversation.
Stack these with Whisper STT and you get voice-in, voice-out AI that responds before you've finished processing that it isn't a person.
Sesame CSM-1B generates natural speech with "umms", "uhhs", and those micro-pauses that make you forget you're listening to a machine. Dia2 produces multi-speaker dialogue with distinct voices, emotion, and timing. CosyVoice2 does streaming synthesis at 150ms first-chunk latency - fast enough for real-time conversation.
Stack these with Whisper STT and you get voice-in, voice-out AI that responds before you've finished processing that it isn't a person.
4
Autonomous browser agent
4090
Browser Use gives an LLM a real browser. Not a simulated one. A real Chromium instance with DOM access, screenshots, and the ability to click, type, scroll, and navigate. It hits 89.1% success rate on the WebVoyager benchmark - a standardized test of web task completion across real websites.
Wire it into n8n and you get automated web tasks that handle the sites with no API. Fill out that form. Check that tracking number. Download that report. The agent reads the page, figures out the interface, and does the thing.
Wire it into n8n and you get automated web tasks that handle the sites with no API. Fill out that form. Check that tracking number. Download that report. The agent reads the page, figures out the interface, and does the thing.
5
Voice command center (Jarvis mode)
4090 CPU
CAAL auto-discovers your n8n workflows via MCP (Model Context Protocol). Every workflow you build becomes a tool the AI can call. Say "check my email" and it triggers the email workflow. Say "back up the NAS" and it triggers the backup workflow.
The key insight: every new n8n workflow is automatically a new voice command. No re-training. No intent classification. No wake-word configuration. You build the workflow, the AI discovers it, and your voice becomes the trigger. The more you automate, the more capable the voice assistant becomes - with zero additional config.
The key insight: every new n8n workflow is automatically a new voice command. No re-training. No intent classification. No wake-word configuration. You build the workflow, the AI discovers it, and your voice becomes the trigger. The more you automate, the more capable the voice assistant becomes - with zero additional config.
Impressive
Not impossible, but deeply satisfying to run on your own silicon.
6
Full-song music generation
4070
ACE-Step 1.5 generates complete songs - verse, chorus, bridge, with lyrics - and needs less than 4GB VRAM. It runs comfortably on the 4070, leaving the 4090 free for other work. At 34x real-time generation speed, a 3-minute song takes about 5 seconds. Benchmarks put its output quality ahead of Suno v5 on multiple evaluation metrics.
7
Voice cloning in 3 seconds
4070
Qwen3-TTS takes a 3-second audio sample and clones the voice with 0.789 speaker similarity score. It handles 10 languages and does zero-shot cloning - no fine-tuning needed. Record a voice memo on your phone, feed it in, and the model speaks new text in that voice.
8
Text-to-3D model pipeline
4090
TRELLIS.2 turns a text prompt into a 3D model in 17 seconds. Not a blob. A textured model with PBR materials, ready to drop into a game engine as a GLB file. The quality gap between "AI-generated 3D" and "artist-modeled 3D" is closing fast, and for prototyping and game jams it's already good enough.
9
3D Gaussian splatting from photos
4090
Take 50-100 photos of a real object or scene. Feed them through a Gaussian splatting pipeline. Get a photorealistic 3D representation you can fly through at 100+ FPS. It captures lighting, reflections, and transparency in ways mesh-based 3D never could. Walk around your apartment with your phone, and ten minutes later you have a navigable 3D replica on your desktop.
10
Multi-GPU tensor parallelism
4090 4070
ik_llama.cpp splits a model's computation graph across both GPUs. A 70B model that would choke on a single 24GB card runs at 30-40 tok/s split across the 4090 and 4070. The graph splitter optimizes layer placement based on each card's VRAM and compute - the 4090 gets the heavy layers, the 4070 picks up the rest.
11
Speculative decoding acceleration
4090 4070
The 4070 runs a small draft model (7-8B) that proposes tokens. The 4090 verifies batches of proposals against the full model. Since verification is cheaper than generation, this yields a 1.5-3x speedup over naive autoregressive decoding. The 4070 effectively becomes a dedicated speculation accelerator - its entire job is guessing what the big model would say next.
12
AI-powered security cameras
4070
Frigate NVR with TensorRT runs object detection on camera feeds: person, car, animal, package. The 4070 handles multiple streams simultaneously with sub-second detection latency. Combined with Home Assistant, it triggers automations based on what it sees - "person detected in driveway after midnight" sends a notification with a snapshot.
13
Native 4K video generation
4090
LTX 2.3 is a 22B parameter video generation model that outputs native 4K with audio sync. Text-to-video, image-to-video, video-to-video - all locally. The 4090's 24GB is just enough to run it with aggressive memory management. The results aren't Sora-level yet, but for B-roll, social content, and creative projects, they're remarkably usable.
14
Diffusion-based image upscaling
4090
SUPIR doesn't just scale pixels - it hallucinates plausible detail using a diffusion model. Feed it a blurry 480p photo and it reconstructs texture, sharpness, and fine detail that was never in the original. Benchmarks put it on par with Topaz Photo AI, the commercial gold standard, but SUPIR runs locally with no license fee and no upload to anyone's cloud.
Combination Builds: The Multiplier Effect
Individual capabilities are impressive. Combining them is where it gets absurd.
"Digital You"
Fine-tune + Voice clone + STT
Fine-tune a model on your writing, emails, and messages. Clone your voice from a 3-second sample. Add Whisper for speech-to-text input. The result is an AI that sounds like you and thinks like you. It answers questions the way you would. It drafts emails in your voice - literally. It's your stand-in for async communication, and it runs on a box in your office.
"Full Production Studio"
Images + Video + Music + Voice + Upscale
Generate concept art in ComfyUI. Animate selected frames with LTX. Score the sequence with ACE-Step. Add narration with voice cloning. Upscale everything with SUPIR. You just produced a short film - visuals, soundtrack, voiceover, post-processing - on one desktop. The entire pipeline runs locally. No render farm. No subscription. No cloud credits. One machine, every step from script to final export.
"See, Think, Act"
Screen vision + LLM + Browser agent + n8n
A vision model watches your screen. An LLM interprets what it sees. Browser Use acts on the web. n8n orchestrates the workflow. The result is a proactive AI assistant that doesn't wait for instructions - it observes context and takes action. It sees a shipping confirmation in your email, extracts the tracking number, checks the carrier website, and adds the delivery date to your calendar. You didn't ask it to. It just did.
GPU Split Strategy
4090
Heavy compute - 32B+ inference, image generation, video generation, 3D, fine-tuning
4070
Concurrent secondary - music generation, voice pipeline, tab completion, draft models, Frigate NVR
CPU/RAM
Background services - Parakeet STT, Kokoro TTS, Qdrant vector DB, n8n orchestration
The thing that keeps hitting me is the acceleration. Half of this list didn't exist a year ago. The models that did exist needed twice the VRAM. The ones that ran locally were noticeably worse than their cloud counterparts. That gap is closing - in some cases it's already closed.
There's something fundamentally different about AI you own versus AI you rent. Not just the privacy angle, though that matters. It's the composability. When every model runs on your machine, you can pipe them together in ways no cloud API would ever allow. Voice into LLM into browser agent into n8n into voice back out. The latency is local. The data never leaves your network. The only limit is your imagination and your power bill.
A $3,000 desktop. Consumer parts. No enterprise hardware. No cloud. Just a PC under a desk doing things that would have required a research lab budget three years ago.
All tools mentioned are open-source or source-available. Benchmarks cited are from their respective project documentation as of March 2026. Your mileage may vary - but probably not by much.