Everything below is measured in our own lab: one model, one GPU, same prompts, fresh chat per cell. Live tables: results.html. Method and numbers: mds/benchmarks.md.
You send a mode tag as plain text with your prompt, for example {REASON:logic} your question. The model has been fine tuned to recognize these tags and change how it answers. No app changes, no API switches, any chat UI or HTTP call works. Two families exist: reasoning modes ({REASON:omni}) and instruct modes ({REASON:iomni}), plus {REASON:off} to return to stock behavior and {REASON:help} to ask the model itself for the menu.
Average completion tokens per mode across our 21 prompt rows (one model: LFM2.5-8B-A1B at Q6_K, RTX 4070). Token count is your money and your latency.
| tier | mode | avg tokens | what it does | pick it for |
|---|---|---|---|---|
| quick | medium-low | 977 | speed first, structured but short | warm ups, drafts, chat |
| low | 987 | minimal tokens | instant factual answers | |
| socrates | 1069 | probes your premises with questions first | critical thinking, spotting bad questions | |
| medium | 1117 | efficient professional structure | quick pro answers, the instruct default | |
| focused | logic | 1262 | first principles, axioms first | debugging, verification |
| high | 1512 | thorough and balanced | the professional default | |
| einstein | 1548 | 20 perspective agents brainstorm | ideation, novelty | |
| hyper | 1851 | MECE exhaustive breakdown | audits, segmentation | |
| ultra | 2046 | perfection chase | high stakes deliverables | |
| structural | spoon | 2531 | 5 expert panel research | research synthesis |
| deeptree | 2536 | recursive tree decomposition | root cause, planning | |
| omni | 2632 | multi dimension holistic pass | strategy, multi variable problems |
Same 10 rows x 12 modes, three quants, same GPU:
| quant | file size | deeptree behavior | reading |
|---|---|---|---|
| IQ4_XS | 4.27 GB | avg 2267 tok, truncated 7 of 10 | cheap disk, chains lose control |
| Q6_K | 6.96 GB | avg 1532 tok, truncated 3 of 10 | the balanced pick, leanest quick tiers |
| Q8_0 | 9.01 GB | avg 956 tok, truncated 1 of 10 | most disciplined, finishes what it starts |
The card rule "quant level gates mode power" shows up in our data as chain control: the lower quant rambles and truncates in structural modes, the higher quant finishes.
{REASON:ilow} if you like, but expect reasoning-grade output.| you want... | send | and expect |
|---|---|---|
| an instant answer | {REASON:low} | short, correct on factual stuff, no structure |
| a solid professional answer | {REASON:high} | balanced thoroughness, 1.5x the tokens of low |
| a plan or root cause tree | {REASON:deeptree} | deep hierarchy, needs a token budget |
| research you can trust | {REASON:spoon} | panel voices, longest chains |
| the premises checked | {REASON:socrates} | questions first, catches bait questions |
| exhaustive coverage | {REASON:hyper} | MECE trees, medium cost |
| a decision across dimensions | {REASON:omni} | weighted verdict, priciest tier |
| a small model on a small GPU | 2.6B Q8 + {REASON:medium} | lean answers, less control in deep modes |
Every claim here reproduces from the repo: docker compose -f docker/compose.yaml up -d llama-tb phoenix lab, then open http://127.0.0.1:3030 and watch the matrix fill in live. Full reproduce steps in mds/benchmarks.md.
Model family: DavidAU Turbo Brilliance (Apache 2.0 weights, his design). Base model: LiquidAI LFM2.5-8B-A1B. We neither own nor maintain the models; credit and support go to DavidAU.