Qwen 3.8 27B on a single RTX 4090, measured (llama.cpp)

An open-source project (awesome-local-ai) ships one-command installers that pair a model with hardware and a stack, and records every performance and memory figure on real hardware rather than estimating it.

Measured reference combination

• Hardware: a single RTX 4090 (24GB)
• Stack: Qwen 3.8 27B served by llama.cpp to OpenCode, at a 128k context
• Speed: ~92 tok/s generation with speculative decoding on, ~44 with it off; the installer asserts the draft acceptance appears in the log before calling the install good
• Vision: a native vision tower is available but off by default; loading it costs 32k of context on a 24GB card (128k down to 96k), a VRAM trade rather than a missing capability

The same project runs a build benchmark (Vidi) that drives a coding agent through a real 17-story specification and scores the result with a held-out test suite the agent cannot read. Setup, hardware, stack and the benchmark methodology are documented in the repository.