Lean Mode: the library that weighs less
Every primitive in the FrootAI catalogue now ships twice. There's the Full version — the verbatim source you've always had — and a new Low-Calorie (Lean) version sitting right next to it as a committed .lean.md file. Flip one toggle and your context window stops paying for whitespace.
That's the whole pitch: Low-Calorie Mode (Lean) — fewer tokens, identical capability.
But I want to be honest with you about the number, because most "compression" stories aren't.
The honest number
Across the entire catalogue, measured with the exact o200k_base tokenizer, Lean takes the load from 1,518,320 tokens down to 1,510,081. That's 8,239 tokens saved — about 0.54%.
Half a percent. Not 50%. Not 70%. Half a percent.
If you came here expecting a dramatic before/after, that's the honest floor, and we publish it loudly on the /lean benchmark page. The deterministic, lossless transform only reclaims trailing whitespace and collapses dead blank lines. It doesn't rewrite, summarize, or drop anything. So the savings are real, reproducible — and modest.
The bigger 50–70% wins people imagine come from semantic compression, and that's a future tier (Lean+). We're not going to dress it up as today's number.
Then why bother?
Because Lean is the only way to make a payload smaller that loses nothing.
Look at the alternatives:
- Truncate — chop the text when the window fills. Huge "savings," but it silently drops whatever fell off the end: a guardrail, half a code block. - Summarize — ask a model to tie it into something shorter. Often 50%+ smaller, but it paraphrases. Exact guardrails and parameters get reworded or dropped, it can hallucinate, and it costs an extra LLM round-trip every time. - Buy more context — skip compression, pay for a bigger window. No fidelity loss, but every token still costs money.
Lean is the row that loses nothing. No round-trip. No paraphrase. No dropped rule. Deterministic and free.
The moat is the fidelity gate
Anyone can make a file smaller. The hard part is making it smaller without changing what it means. So before any Lean variant ships, it has to clear a fidelity gate:
- Guardrails are preserved verbatim — Lean can never quietly relax a safety rule. - Parameters, defaults, and enums are kept exactly — the model gets the same instructions to act on. - Code blocks are never reflowed inside the fence — what compiles in Full compiles identically in Lean.
The verdict across the catalogue is intentionally flat: 964 of 964 Lean variants scored a perfect 10/10. That's not a marketing round-up — a variant that drops a guardrail, parameter, or code block simply never makes it into the catalogue. That's the moat: lossless by construction.
Turning it on
Lean rides the same FROOTAI_LEAN signal and .lean.md artifact everywhere:
- CLI — frootai install , or compile your own with frootai lean .
- SDKs — leanMode: true (npm) / lean_mode=True (Python).
- Runtime — set FROOTAI_LEAN=true and the GitHub Action, the Foundry agent, the Docker image, and the Helm chart all prefer Lean.
Small and honest beats big and lossy
When you actually need fidelity — and with guardrails and parameters, you always do — a guaranteed 0.54% beats a hopeful 50%. Meet Lean Mode at frootai.dev/lean-mode, and see every measured figure on the benchmark.
Same capability. Fewer tokens. Nothing given up.