Running agents at home is something most people can afford to do, and something that will improve their lives, often on hardware they already own. A year ago this was not the case. Local models were the domain of large businesses, and highly technical tinkerers.

• The models got smarter.
• The software got faster.
• The hardware got more expensive.
• The agents got useful.

We’ve combed 184 websites and collected Hardware prices and sales data, model configs, and hardware registries to put together a high level overview of the industry, community, and technology

Why now

Key takeaway: a model that fits one 24 GB GPU now scores where the best model in the world was in February 2026. It came out about six months after that frontier model.

• The same size got much smarter. Qwen3.8 27B scores 33.7 on the Artificial Analysis Intelligence Index. The Qwen3 32B it replaced scores 8.6 on the same version of the index. It also needs a quarter of the KV cache (the memory of everything already in the conversation) per token and handles 8x the context.

Qwen's dense 27B models and 33B parameters (dots) against every new frontier record (lab logos), all scored on the same version of the Artificial Analysis index. The dashed line is the time from GPT-5.3 Codex to Qwen3.8 27B at the same score. Hollow dots are scores AA estimated. Source: Artificial Analysis, September 2026 snapshot.

• Software doubled some setups for free. On my own RTX 3090, moving Qwen3.8 27B from AWQ 4-bit to EXL3 3-bit (two compression formats, covered in section 2) with multi-token prediction, in the same engine, took decode from 45 to 99 tokens per second. Ollama reported +90% on an M5 Max from multi-token prediction for Gemma 4, and LM Studio's MLX engine doubled speed on Apple Silicon.

• Mixture-of-experts fits the hardware people buy. A mixture-of-experts (MoE) model only reads the few experts it routes each token to, so it runs much faster than a dense model of the same size. In every matched pair in the registry, the MoE won.
• Renting got less comfortable. When Cursor moved to usage-based pricing I started worrying I might not always be able to afford the tools I work with, and that is how my homelab started.
• Hardware got more expensive. Most GPUs sell well above launch price, NVIDIA raised the DGX Spark from $3,999 to $4,699 in February and in-stock Sparks now sell for about $5,700, and Apple doubled its memory-upgrade prices in June. The cheapest way to hold a big model today is a 128 GB AMD Strix Halo, at $30 per gigabyte a model can use.
• More people are doing it anyway. Between May and September 2026, the number of people listing each popular device on their Hugging Face profile grew 2.6 to 8.3 times, and new MLX model builds per month nearly tripled since January.

1. Hardware

Key takeaway: memory size decides what you can run, memory bandwidth decides how fast. Big unified-memory boxes hold the most model per dollar; discrete GPUs are the most speed per dollar.

Two numbers decide almost everything