Startup Mirendil claims a Kimi K3-level model trained by one researcher for under $10M
Ok_Warthog6007 · reddit · 2026-11-09
Startup Mirendil claims it trained a model at Kimi K3 level using just one researcher, 3.5 months, and under $10M in compute, detailing the effort in a blog post titled "Pretraining at Scale Without a Pretraining Team."
If accurate, this suggests frontier-level pretraining cost thresholds are collapsing for tiny teams. The claim is self-reported and not yet independently verified, so the "K3-level" benchmark framing deserves skepticism.
More from Venture
- AI selloff prompts rethink of Anthropic's expected valuation — pstAsiatech · 2026-11-09
- Scoble backs NVIDIA staff through rough week: 'good chance you will have a new leader' — Scobleizer · 2026-11-09
- The funniest part of the Jensen Huang stock dump: the news was 24 hours old — econoar · 2026-11-09
- Could coding agents make the FDE deployment motion viable below enterprise tier? — curious_vii · 2026-11-09
- AI consultant's business model: arbitraging the gap between the center and the frontier — curious_vii · 2026-11-09
- Short seller Culper alleges Nvidia has a 'China problem'; GPU shipment claims under scrutiny — bookwormengr · 2026-11-09