How we know
Observations
Most numbers on this site come from someone else's page. These we measured or recorded ourselves, on one Mac mini with an Apple M1 and 8 GB of memory. Each says how it was taken and what it does not show.
Qwen3 next-token probabilities
Recorded 11 November 2026 by Claude. Used by Plate V.
Ollama 0.32.5 served the model tag qwen3:1.7b (Q4_K_M). Each of four prefixes was sent once as plain text with no chat template, asking for one token with logprobs on andtop_logprobs: 20. The page stores the 20 returned log probabilities for each prefix, rounded to five decimals.
On this mini on 11 November 2026, ollama show qwen3:1.7b reported temperature 0.6, top-k 20 and top-p 0.95.
Limits: one run per prefix, one small model. Ollama's metadata labels the installed file 2.0B parameters, despite the tag's 1.7b.
Local generation speed
Measured 11 November 2026 by Codex. Used by Plate V and Plate VI.
Five warm requests to the same tag, each writing 256 tokens with thinking off, at temperature 0 and a 2,048-token window. Speed is the server's own count: tokens generated divided by generation time, which leaves out loading and reading the prompt.
Five warm requests had a median output-generation rate of 36.87 tokens/s, with a range of 36.56 to 37.11. The model file was 1,359,293,444 bytes; once loaded, Ollama reported 1,530,627,357 bytes in use.
Limits: other agents and an Android emulator were running, and the machine was swapping. One prompt, one model, one context size. Time to the first token was not measured.
Sending agents from a phone
Recorded 17 November 2026 by Claude, from the studio's own decision records. Used by Personal Agents.
Micah Marino, who runs this site, built his own way to start agents. It is an installable web page on his phone with one record button. The recording is kept on the phone until his Mac mini at home confirms it has the file, and it travels over his private network, so nothing is open to the internet.
On the Mac mini, a local whisper model writes the transcript word for word. Each capture is saved as a Markdown file and its audio in a git repository, labelled idea, project or action.
An action capture starts Claude Code in a new Herdr workspace. That steward agent reads his words and starts work only if they ask for it now. It then starts a new project with its own repository, workspace and agent, or hands the request to the agent already working on that project. The build agents may use subagents and worktrees.
Spending money, public releases and deleting his work still wait for him. On 15 November 2026 a capture that said "I want to start this in a new session right now" became a new project with its own workspace and agent.
Limits: one person's setup on one Mac mini, working only while it is on. The agents run with his own permissions. It reads no email, messages or bank accounts; it builds software.