The cloud AI companies have trained us to think in subscriptions, usage caps, rate limits, and tokens that behave suspiciously like casino chips.

You can opt out of a surprising amount of that right now.

For roughly $2,500 to $4,700, you can put a machine on your desk, download serious open-weight AI models, and run them as much as your electricity bill and hardware can tolerate. Your prompts and files can stay on your machine. There is no per-token API charge. There is no vendor deciding you used too much AI this afternoon.

One clarification matters: "unlimited tokens" means unlimited local usage, not infinite context. Every model still has a context-window limit, and larger contexts consume more memory. OpenAI's gpt-oss models, for example, currently offer a 128K context window.

Heroik’s recommendation: If you and your org can swing it, find the be

That said, here are 3 rigs I would look at today.

1. The Easy Mac: Mac mini M4 Pro, 48GB

Current price: about $2,499 from Apple for the 48GB/1TB configuration.

This is the easiest path for someone who wants local AI without turning local AI into a new hobby. Apple's M4 Pro provides 273GB/s of memory bandwidth, and Apple Silicon's unified-memory architecture gives local models access to a large shared memory pool rather than forcing everything through a conventional dedicated GPU.

Load macOS, install LM Studio, and start with OpenAI's gpt-oss-20b. The current Ollama package is about 14GB, so 48GB gives you meaningful room for context and other applications.

A llama.cpp benchmark on a 64GB M4 Pro Mac mini measured about 63 tokens per second for gpt-oss-20b. Different runtimes and settings can move that number substantially, but 55 to 65 tokens per second is a sensible expectation for a straightforward setup.

That is genuinely fast. Artificial Analysis currently measures GPT-5.6 Sol Max at roughly 63.5 output tokens per second through OpenAI's API, although Amazon's hosted implementation can exceed 100.

Do not confuse similar output speed with similar intelligence. GPT-5.6 Sol is a substantially more capable frontier model. The comparison tells us something else: local inference no longer has to feel painfully slow.

Buy: Apple Mac miniApp: LM StudioModel: gpt-oss-20b

Setup is almost boring: install LM Studio, open Discover, search openai/gpt-oss-20b, download the recommended local version, load it, and chat. LM Studio can also expose the model as an OpenAI-compatible or Anthropic-compatible local API, letting other software use your Mac as its AI engine.

2. The Value Monster: Framework Desktop, Ryzen AI Max+ 395, 128GB

Current refurbished price: $2,929 for the 128GB version. New-system pricing moves with configuration and today's ugly memory market.

This is where things get interesting.

The Framework Desktop uses AMD's Ryzen AI Max+ 395 with Radeon 8060S graphics and 128GB of LPDDR5x unified memory. Framework allows up to 96GB of that pool to operate as GPU-equivalent memory. That is enough to run OpenAI's much larger gpt-oss-120b locally.

Framework publishes an official benchmark of 38 tokens per second for gpt-oss-120b MXFP4 using LM Studio on Fedora 42.

For perspective, that is roughly 60% of the current OpenAI-hosted GPT-5.6 Sol Max output rate, while a 120B open-weight reasoning model is running under your own roof. Again, the frontier cloud model is smarter. But 38 tokens per second is comfortably conversational.

Buy: Framework DesktopOS: Fedora LinuxApp: LM StudioModel: gpt-oss-120b MXFP4

The build path is straightforward. Install Fedora, install LM Studio, open Discover, download gpt-oss-120b, then load it. Framework specifically publishes its performance numbers using LM Studio on Fedora, so I would resist the urge to get clever until the machine is working.

For someone willing to operate Linux, this may be the most interesting current combination of big-model capability, memory, size, and price.

3. The AI Workstation: NVIDIA DGX Spark, 128GB

Current NVIDIA price: $4,699. NVIDIA increased the original $3,999 MSRP in February 2026 because of memory-supply constraints.

The NVIDIA DGX Spark is the least DIY computer here and the most purpose-built AI machine. It uses NVIDIA's GB10 Grace Blackwell platform, 128GB of coherent unified memory, a 4TB NVMe drive, and NVIDIA's DGX OS. NVIDIA supports models up to 200B parameters on a single Spark, with 2 Sparks linkable for models up to 405B.

For local text generation, gpt-oss-120b is the obvious starting point.

Current optimized llama.cpp testing has produced roughly 55 to 59 tokens per second from gpt-oss-120b on the Spark. Ollama's earlier standardized test measured 41.14 tokens per second, which is a useful reminder that firmware, runtime, context length, and optimization matter.

Buy: NVIDIA DGX SparkOS: NVIDIA DGX OSApp: OllamaModel: gpt-oss-120b

Installation is wonderfully short. Install Ollama, then run:

ollama run gpt-oss:120b

That command downloads and launches the model locally. Ollama can also serve your model as a local API and connect local models to development tools including Codex.

Which One Should You Build?

For most people, I would start with the Mac mini if simplicity matters, the 128GB Framework if model size per dollar matters, and the DGX Spark if you want NVIDIA's AI ecosystem and expect to experiment heavily with agents, development tools, fine-tuning, or multi-machine setups.

The bigger point is not which box wins a benchmark.

A useful local AI capability is now obtainable for the price of a serious workstation. You can own the hardware, hold the data locally, choose the model, replace the model when something better appears, and generate millions of tokens without watching a usage meter spin.

Cloud AI still owns the absolute frontier.

Local AI is beginning to own something else: control.

For operators building proprietary knowledge bases, company brains, private research environments, coding systems, and always-on agents, that is a very different kind of leverage.