Local AI
Run open-weight models on your own hardware: what fits, what it costs, and what it runs like.
-
Strata: the Engine That Made 125B a Gaming-PC Hobby
Open-source local inference system for Windows and Linux: experts into RAM, 29 GB lookup table on SSD, GPU handles only the active slice. What it runs, at what size, at what speed.
-
Qwen3.8-Flash-Next: the 125B That Made 100B-at-Home a Hobby
The face of the trend: 125B total, multimodal, and packed small enough that a 48 GB PC + any recent GPU runs it at 76–94 tok/s. Exact pack sizes, measured speeds, and which variant to pick.
-
Project Maya: the Scheduler That Puts 748B on Your Desk
Open-source (MIT) three-tier engine: hot experts on the GPU, next ones in RAM, the rest on NVMe — moving as you chat. Runs GLM-5.3-Flash, MiniMax-M3.1-Flash and DeepSeek-V4.1-Flash-1017 on hardware you already own.
-
MiniMax M3.1-Flash: the 428B 'Space Bunny' Everyone Switched To in October
Open weights October 23, 2026: 428B total, 23B active, native multimodal, 1M context, ~70% cheaper to serve than M3 — the newest member of the home-rack trio.
-
How a 100B+ Model Fits in a Gaming PC — what Strata and Project Maya changed
The 2026 local-AI wave is one trick: MoE experts live in RAM, the lookup table lives on SSD, the GPU only touches the active slice. Here's the trick, the numbers, and which engine to run.
-
GLM-5.3-Flash: the 320B That Fits Because Its Attention Does Not Grow
Zhipu's 320B-total, 18B-active MIT-licensed multimodal MoE is the agentic flagship of the home-AI wave — hybrid linear+sparse attention cut compute ~3× and KV cache to a third, which is exactly why 64 GB boxes can run it.