Local LLMs in 2026 closed enough of the gap with cloud frontier models to matter. This hub tracks what runs where, how fast according to published measurements, and where the bottleneck still sits between you and Claude-tier capability.
A Mac mini M4 Pro with 48 or 64 GB of unified memory loads Llama 3.3 70B at Q4, a 43 GB file in the Ollama library. On llama.cpp's published Apple Silicon scoreboard the same chip writes about 50 tokens per second with a 7B model; no published scoreboard gives a 70B speed. That is the ceiling of local models at small form factor today.
The economics flipped too. A €1,899 Mac Mini paid back over 24 months is around €88 per month all in. A heavy Claude Sonnet user can burn that in API calls a few weeks. The break-even is real for sustained agentic workloads, even if cloud frontier remains better for the hardest single-turn tasks.
This hub gathers the published speed measurements, the hardware that runs each model class, the tooling (Ollama, llama.cpp, GGUF), and the agent integrations that bring local models into self-hosted setups. Start with the Local LLM benchmarks page if you want numbers, the hardware buyer's guide if you want a buying decision, or ZeroClaw / Nanobot if you want agents that prefer local.
Guides
- Local LLM benchmarks on a Pi 5, mini PC, Mac or GPU: Published llama.cpp speeds for one 7B model on mini PCs, Apple Silicon Macs and Nvidia cards, Raspberry Pi 5 figures, and the memory each model size needs.
- Pocket AI: which self-hosted agents run on small hardware: Which self-hosted AI agents run on a Raspberry Pi, mini PC, Mac mini, Framework laptop or old phone: an agent-by-device matrix, the limits, six setups.
- Best hardware for local LLMs and AI agents in 2026: Raspberry Pi 5, mini PC, Mac mini or Mac Studio for a local LLM or a self-hosted AI agent: what each budget buys, with makers' specs and published speeds.
Agents on this topic
- ZeroClaw: Single Rust binary from ZeroClaw Labs, MIT or Apache-2.0. Works with Anthropic, OpenAI, Ollama and about twenty other providers, and turns on an OS sandbox it detects.
- Nanobot: Personal AI agent in Python (MIT) with a browser UI and chat apps. Native Anthropic, OpenAI and OpenRouter providers plus local servers; its shell sandbox is off until you switch it on.
- Hermes Agent: Agent from Nous Research (MIT, Python). Runs shell commands on the host by default, with an approval step for dangerous ones; Docker and cloud sandboxes are options. Works with many LLM providers.
Hardware on this topic
- Mac Mini M4 / M4 Pro: Mac mini M4 and M4 Pro for local LLMs: 16 to 64 GB of unified memory at 120 or 273 GB/s, the two specs that decide which models load and how fast.
- Mac Studio M3 Ultra: Apple's 2025 local-LLM workstation: M3 Ultra, 96 GB of unified memory configurable to 256 GB, 819 GB/s. Since replaced by the M5 Ultra model.
- Minisforum UM790 Pro: Minisforum UM790 Pro: Ryzen 9 7940HS with Radeon 780M graphics and up to 64 GB of DDR5, for a Linux agent host that also runs local models.
- Intel NUC 13 and Geekom IT13: Intel NUC 13 and Geekom IT13: 13th-gen Intel Core mini PCs with 16 to 64 GB of RAM, room for Hermes Agent with browser automation and a 7B model beside it.
- Raspberry Pi 5: Raspberry Pi 5 for self-hosted AI: four Cortex-A76 cores, 1 to 16 GB of LPDDR4X. Hosts Hermes Agent (no browser tool) or Nanobot; local models up to about 3B.
Terms
Local LLM · Llama · Mistral · Qwen · Phi-3 · Quantisation · Ollama · llama.cpp · GGUF