Nvidia Backed Reflection AI Struggles to Ship Amid Open Source Race

The New York lab was originally building an AI coding tool, and per the reporting, the co-founders are "completely 100% steering the company away from their vision to Jensen's vision." The catch-up framing is theirs.

What Nvidia has put behind it

This isn't a quiet pivot. According to the same reporting, Nvidia's commitment runs roughly: $800M to start, 40% of Reflection's $2B round in October, and more this year at a $25B valuation. In March, the company formed the Nemotron Coalition — eight labs pooling research, data, and compute. On July 24, Nvidia, Meta, Microsoft, and 22 other organizations signed a joint statement backing open-weight models; Huang's first-ever X post was about that statement.

The open-weight scoreboard

Here's the uncomfortable part for anyone tracking releases. According to AI: Reset to Zero's breakdown, the four leading open models are Chinese. The two American entries sit at the bottom of that list, and one of them is Nvidia's own Nemotron 3 Ultra.

Concrete numbers from the same write-up: Moonshot's Kimi K3 shipped this month at 2.8 trillion parameters — the largest open-weight model released so far. Alibaba previewed Qwen3.8-Max days later at 2.4 trillion parameters, with open weights promised but not yet out. Thinking Machines' Inkling sits just under a trillion. Nemotron 3 Ultra is 550 billion.

What to benchmark against while we wait

Reflection's founder and CEO Misha Laskin told CNBC in April that "these things are hard to build. They're kind of like rocket ships." Fair — rockets are hard. But it also means we shouldn't paper over the gap with a name drop in the architecture diagram.

Practical sanity check: pin baselines against the open models that actually shipped — Kimi K3, the Qwen3.8-Max preview, Inkling, Nemotron 3 Ultra — rather than against the ones announced. Treat any "coming soon" Reflection line item as TBD. Watch the Nemotron Coalition output; that's where the pooled compute actually surfaces as weights.

The size class is shifting. Anything we benchmark at 70B today sits in a different conversation than the 2.4–2.8T range the Chinese labs are previewing. Plan scaling tests around the new envelope, not around last quarter's leaderboard.