A collection of 10 posts
H Company's Holo3.1 family brings computer-use agents to local and on-device inference with quantized checkpoints and four model sizes. Here's what shipped and how to deploy it.
May 2026 was a heavy ship month for local AI runtimes. Ollama added Codex App support. vLLM 0.21 stabilised DeepSeek V4 on Blackwell. llama.cpp merged MTP speculative decoding. MLX hit 4x faster on M5. LM Studio shipped stable MTP. Practical runtime-by-runtime changelog.
Honest 2026 comparison of the five dominant local LLM runtimes: Ollama, LM Studio, vLLM, llama.cpp, and MLX. Throughput numbers, feature matrix, and a decision tree.
Quick answer. To run Gemma 4 on Windows, install Ollama from ollama.com, open PowerShell, and run ollama pull gemma4:e4b followed by ollama run gemma4:e4b. The E4B (~9.6 GB) variant fits comfortably on 16 GB systems. For a GUI, install LM Studio, search “gemma 4”, download a
Eight free local LLM runners ranked for 2026, with a decision tree, VRAM math for 7B to 200B models, and notes on the Apple Silicon and open-weight wave that made laptop-class frontier AI real.
Qwen 3.7 weights are not on Hugging Face yet (May 20, 2026). Here are the honest ways to use it today, and exactly what to run locally instead.
What OmniCoder 9B is, its lineage and license, vendor-reported benchmarks, the full GGUF quant table, and step-by-step Ollama and llama.cpp setup.
Void is a free, open-source, VS Code-based AI code editor and Cursor alternative. Here's what it does, how it works, and whether the paused project is worth using in 2026.
Run Qwen 3.6 locally: 27B dense vs 35B-A3B MoE explained, VRAM tables per quant, and copy-paste Ollama, llama.cpp, vLLM, and MLX commands.
A 2026 decision framework for vLLM, Ollama, and LM Studio — when each one wins on throughput, hardware support, and cost, with cited benchmarks instead of fabricated numbers.