Nicheloom visits: 29,129 all-time · 11,440 in the last 7 days

The opportunity tracker for new startups.

Today: 1,144 launches · Niche of the day: consumer buying and practical guides · See full report →

reki

REKI: a 2.4B vision-language-action model pretrained on public data only, with a memory of its recent past at negligible cost.

Get picks like this daily. Today's launch numbers, the top launches, AI and tech news, and a daily opportunity pick, straight to your inbox.

Visit website ↗ Source listing ↗

GITHUB · launched Oct 15, 2026 · 24 upvotes (top 50%)

Similar launches

See all 2049 similar launches →

Could you build this?

No REKI is a research-grade 2.4B vision-language-action model whose contribution depends on large-scale pretraining on curated robot and human data and GPU training runs. Vibe coding can't produce the model or the novel memory architecture.

What it would actually take: Assemble a VLM backbone with a flow-matching action expert and a register-token memory design, then pretrain on a curated mixture of public robot, egocentric, and vision-language datasets. The real cost is compute (many GPU-days) plus ML research expertise for architecture design, data curation, and evaluation on LIBERO, RoboTwin, and real hardware such as a UR5 arm.

Competitors

Other products that read as similar to this one : 2049 launches clear the similarity bar, closest 8 shown.

Attention rank: #1020 of 2050 (itself plus its competitors, highest first, normalized so YC and Product Hunt are compared fairly).

Launched 350 days after the earliest competitor.

Other launches for this product

More details
External ID
1421011116
Company
n/a
Product
reki
Website domain
kyutai.org
Cohort
n/a
Upvotes percentile
0.50
All tags
n/a
AI summary
vision-language-action model with past memory for robotics and embodied ai
Manually corrected
No