We gave nine setups the same seven AI-written drafts and asked for three rewrites of each. The model trained for the job was the fastest by far and reworded most of the quotations its own prompt told it to keep. General models following the open humanizer skill kept nearly everything, including a 27B running on the same GPU.
PrismML compressed Qwen3.8-27B to ternary weights and kept 98.2% of its benchmark average by its own measurement. The weights shrank nine times. The KV cache did not move, and the capability that degraded most is the one we build on.
Alibaba shipped a 27B model in August that scores 52 on the Artificial Analysis index against its predecessor's 38, with an identical network. What that means for running a capable model on one graphics card, and what it costs in memory and tokens.
The top four models on the Artificial Analysis index sit within four points of each other and vary threefold in cost per task. The cheapest of them is the one you can download. What changed in July was pricing power, not capability.
NVIDIA, Microsoft, Meta, and 22 others signed a letter arguing the US should not restrict downloadable models. We read the three pages, checked the one claim we can verify from our own work, and noted what the letter avoids.
The hardware to hold a capable model got cheap, and the models got small enough to fit. What runs on the machines in your bedroom, and the two numbers that decide it.
Anthropic's engineering blog describes how to build agents in production. agent-stdlib packages those procedures into installable skills, MCP servers, and safety hooks you can run in any harness.
This website is a public GitHub repository, and so is the rest of our stack. The practical case for building on open source.
We built our production website with Qwen3.6-27B, an open-source model from Alibaba. What frontier open-weight models can do now, and where you still need a person in the loop.