- L1–10Script
- L11–20Function
- L21–30Module
- L31–40Package
- L41–50Library
- L51–60Framework
- L61–70Runtime
- L71–80Compiler
- L81–90Kernel
- L91–100Architect
LEVELS
How levels work
Your level decides how many requests may run at the same time. It rises with the tokens you spend on software work and falls when you use devseek for other things. Each step costs more XP than the last.
The ladder
Every 10 levels
| Level | Rank | Concurrent sessions | Total XP to reach | XP to next level |
|---|---|---|---|---|
| L1 | Script | 2 | 0 | 692 |
| L2 | Script | 2 | 692 | 1,407 |
| L10 | Script | 2 | 23,291 | 4,277 |
| L20 | Function | 3 | 76,986 | 6,585 |
| L30 | Module | 4 | 151,441 | 8,442 |
| L40 | Package | 5 | 243,282 | 10,058 |
| L50 | Library | 5 | 350,527 | 11,515 |
| L60 | Framework | 6 | 471,814 | 12,860 |
| L70 | Runtime | 7 | 606,130 | 14,116 |
| L80 | Compiler | 8 | 752,681 | 15,302 |
| L90 | Kernel | 9 | 910,818 | 16,429 |
| L100 | Architect | 10 | 1,080,000 | — |
Ranks
Every ten levels earn a new rank and mark.
- ScriptL1–10
- FunctionL11–20
- ModuleL21–30
- PackageL31–40
- LibraryL41–50
- FrameworkL51–60
- RuntimeL61–70
- CompilerL71–80
- KernelL81–90
- ArchitectL91–100
The rules, exactly
- Every request that finishes is classified once as coding or not coding. Requests in the same conversation reuse the verdict for up to 20 requests.
- Coding: +1 XP per 750 tokens. Not coding: −1 XP per 1,500 tokens. Tokens are the prompt and completion counts reported by the model, never what the client claims. You gain at most 90,000 and lose at most 45,000 XP per day (UTC).
- Once a day XP is compared with the ladder. You go up to every level your XP has reached, several at once if you earned enough. If XP falls below 85% of your current level's XP you go down one level, never more than one a day.
- On a day with no coding requests XP is multiplied by 0.97.
- Level 1 is the floor and level 100 the cap. An operator can set a level by hand; every such change is recorded.
- Total XP to reach level L is round(1,080,000 × ((L−1)/99)^1.6). Sessions at level L are 2 + floor((L−1)×8/99).
- Each coding request is also graded as a training record. XP multiplier: good ×5, normal ×1, weak ×0, junk ×-1 (junk takes XP away, within the daily loss limit). Weak and junk records are deleted, not stored. A reply where the model showed no text at all is not your fault: it counts as normal and is not stored. The ladder is 20× the original one; XP per token and the daily limits are set so that a heavy daily coder can reach level 100 in about three weeks.
How a request is classified
Clear cases are decided by local rules: code fences, stack traces, diffs, shell commands, file names, tool definitions and programming words. Unclear requests may be sent to jev, a small decision model, as one yes-or-no question built from your request text, up to 240 times a minute for the whole service. If jev is unavailable the local rules decide. The source of every verdict is recorded.