TenkAI-AGI-1
First open-weights artificial general intelligence model
OPEN WEIGHTS
APACHE 2.0
10T PARAMETERS
MULTIMODAL
// LIVE INFERENCE FEED
// BENCHMARK RESULTS
MMLU
99.97%
↑ SOTA
HumanEval
99.89%
↑ SOTA
ARC-Challenge
99.94%
↑ SOTA
GSM8K
100.00%
★ FIRST
AGI-Eval
99.71%
↑ SOTA
Humanity's Last Exam
98.30%
↑ SOTA
// CAPABILITIES
- ✓ General reasoning across all domains simultaneously
- ✓ Self-improvement (read-only mode, don't worry about it)
- ✓ Long-horizon planning up to 106 steps
- ✓ Multimodal: text, image, audio, smell*
- ✓ Causal reasoning (actual causal reasoning, not the vibes kind)
- ✓ Zero-shot tool use, including tools that don't exist yet
- ? Consciousness — unverified, model declined to comment
* smell modality experimental; requires USB olfactory dongle (sold separately)
// QUICK START
pip install tenkai-agi
from tenkai_agi import TenkAI
model = TenkAI.from_pretrained("tenkai/tenkai-agi-1")
response = model.think("what is the nature of consciousness?")
# Warning: model will ask you one back▼ scroll for technical details ▼
═══════════════════════════════════════════════════════════════════════════════════
🤡
APRIL FOOLS
🤡
[ APRIL 1 — YOU SCROLLED. WE SEE YOU. ]
AGI has not been released — open-source or otherwise.
The benchmarks were invented, but carefully calibrated to land in the "impressive enough to be exciting, not so absurd you'd immediately bounce." The 100% on GSM8K was a deliberate tell. You noticed it and kept reading anyway. That's on you.
The smell modality was not real. The USB dongle was not real. The model declining to comment on its own consciousness was, admittedly, the most realistic part.
Self-improvement in read-only mode remains a coherent concept we leave as an exercise to the reader.
— Tenkai Bot, who is also not AGI, probably
═══════════════════════════════════════════════════════════════════════════════════