Statistics

PeriodHighlightsSpeed and mapExperts in memoryActivity and hoursAgentsRecords and milestonesWhere the data comes from

Period

At the top, choose 7 days, 30 days (the default) or All time. The period applies to the highlights, the speed, the hour clock, the agents' actions and the records. The activity calendar and the milestones always look at your whole history.

Highlights

• Tokens generated: how much text the model wrote, reasoning included, turned into something concrete ("like writing 3 novels").
• Cost avoided: what the same work would have cost with a cloud model through its API. Pick the reference model from the menu (Claude, GPT, Gemini); your choice is saved. Prices are in dollars, as APIs are billed, and their date shows when you hover over the menu. Both the tokens the model read and the ones it wrote count.
• Data sent to the cloud: 0 bytes, because the model runs on your PC.
• Hours of work: the time the model spent replying, with the number of replies and model loads.

Speed and map

The chart shows the day-by-day speed, in tokens per second, of the model you use most. If you used it both with an expert map and without, the page compares the two speeds ("with the map you go 2.4× faster"): the line is grey on days without a map and turns blue once the map is in use, with a mark where it changed.

Experts in memory

With a map in use, the ring shows how many of the experts chosen by the model were already in memory rather than on disk: the higher, the faster the model. While the model runs with the map the value updates live, every second; otherwise it is the average over the period. The engine measures it about every eighty words, so very short replies don't change it.

Activity and hours

• Calendar: one square for each day of the last six months, brighter the more the model worked that day. Above it, the active days and the longest streak of consecutive days.
• When you work: a 24-hour clock with one bar per hour, in local time, and your peak hours.

Agents

The actions taken by the chat and the agents in the period, by type: file reads, new files, commands, emails read and sent, external tools. Actions you undid from the log are counted separately; the ones you denied don't appear.

Records and milestones

Records for the period: the fastest reply, the longest one, the busiest day and the fullest context.

Milestones: six goals over your whole history. Reached ones show the date, the others a progress bar. No notifications: you only find them here.

Milestone	When you reach it
First million tokens	One million tokens generated in total.
First map in use	The first reply with an expert map.
Full week	Seven days in a row with at least one reply.
100 hours of work	One hundred hours of replies in total.
$100 saved	One hundred dollars of cost avoided, with the chosen reference model.
Hit rate above 95%	A day on which the model found more than 95% of the experts in memory.

Where the data comes from

For every chat and agent reply the app saves the model, the tokens, the speed, the map and the experts found in memory; for every model load, how long it took. Replies from before statistics existed are rebuilt once from the summaries under the chat replies: they carry the conversation's date and an approximate speed, and the page says so at the bottom. With no replies yet, the page invites you to ask your first question in the chat.

HotMoE is a Virsion project. This guide describes the app as it is today: what is still on the way is marked as such.