ML
Misha Laskin
CEO. Former research scientist at Google DeepMind, where he worked on reinforcement learning and reward modeling for Gemini.
The company
Reflection AI, the lab behind Beam
A New York startup founded in 2024 that went from autonomous coding agents to training one of the largest open-weight models built outside China in under two years.
- Founded
- 2024
- Headquarters
- New York (Brooklyn), USA
- CEO
- Misha Laskin
- Disclosed funding
- More than $4 billion
- Last reported valuation
- $25 billion pre-money
Founders
IA
Ioannis Antonoglou
Co-founder. One of DeepMind's earliest engineers and a co-creator of AlphaGo, AlphaZero and MuZero; later worked on reinforcement learning from human feedback for Gemini.
“The only way to own intelligence is, by definition, if it's open — because if it's open, you can run it on your stuff, you can own it, you can control it, you can customize it.”
From coding agents to open frontier models
Reflection spent its first year in stealth and came out in March 2025 with $130 million and a plan to build superintelligent autonomous coding systems first, then expand to other computer work. Its first product, Asimov, was a code-research agent that helps teams understand large codebases.
In October 2025 it raised $2 billion at an $8 billion valuation in a round led by Nvidia, and repositioned around open frontier models for companies, governments and public institutions that want full control over their AI stack.
Beam, announced on October 5, 2026, is the first model of that new strategy and the first in a planned series. Alongside it Reflection released Mirror CLI, a terminal coding agent that uses Beam by default.
Funding
| Date | Round | Amount | Valuation | Notable investors |
|---|---|---|---|---|
| March 2025 | Seed and Series A | $130M | ≈ $545M | Sequoia Capital, CRV, Lightspeed Venture Partners, NVentures, Databricks, Reid Hoffman |
| October 2025 | Series B | $2B | $8B | Nvidia (lead), Citi, Sequoia, Lightspeed, Eric Schmidt, 1789 Capital, GIC, DST Global, B Capital |
| 2026 | Latest round | ≈ $2.5B | $25Bpre-money | Reported participants include JPMorgan and Disruptive |
Amounts and valuations as reported by Reuters, The Wall Street Journal and Business Insider. Reflection does not publish full financing details.
Compute
Beam's pretraining ran on 6,144 NVIDIA GB300 NVL72 GPUs and its reinforcement learning on about 10,500 GB300 GPUs. In June 2026 Reflection signed a multiyear agreement for GB300 systems at SpaceX's Colossus 2 data center, worth up to $6.3 billion if it runs through 2029.
Why the “DeepSeek of the West”?
In March 2026 The Wall Street Journal reported that some investors were using the phrase. The comparison is about strategy rather than nationality: like DeepSeek, Reflection bets on efficient models with openly released weights. The difference is that it is an American company, which matters to governments and enterprises that do not want to build on Chinese models. Reflection's own benchmarks are modest about the comparison: they show Chinese open models such as Kimi K3 and Qwen 3.8 Max still ahead on raw capability.