Reflection AI, Explained: What the Company Is and What It Has Actually Shipped

Reflection AI is a US artificial intelligence company developing open-weight frontier models. Founded in 2024 by two former Google DeepMind researchers, it has raised billions of dollars from investors, including Nvidia.

Its first model, Beam, was announced on October 5, 2026, and its open weights went public on October 16, 2026, on this page.

This guide separates what has been confirmed from what has merely been reported and what remains unknown. If the company has not verified a number or date, we say so rather than guess.

What Is Reflection AI?

Reflection AI is a private US AI company whose stated mission is to "build the future of open intelligence." On its own site it says it builds "open models that let anyone control their intelligence and help shape the future of AI."

The positioning is explicitly geopolitical. The company presents itself as an American open-weight counterweight to Chinese open-weight leaders such as DeepSeek, which currently dominate the downloadable-model market.

Reflection publishes three commitments on its about page: release open weights for its models, publish the research behind them as papers, and open-source software so organizations can customize deployments.

• Type — private US AI company, founded 2024.
• Stated mission — "Build the future of open intelligence" and "make intelligence open and accessible to all."
• Stated commitments — open weights, published research papers, open-sourced software.
• Named partners on its own site — Dell, the U.S. Department of Energy, and Shinsegae Group.
• Status as of October 16, 2026: Beam announced October 5, weights published October 16 under Apache 2.0.

Who Founded Reflection AI?

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou. Both came from Google DeepMind.

Laskin is co-founder and CEO. Antonoglou is co-founder and is widely credited as a co-creator of AlphaGo, DeepMind's Go-playing system. The AlphaGo project was led by David Silver, who left DeepMind to found Ineffable Intelligence.

The wider team has been described as researchers and engineers who worked on DeepMind projects including Gemini, AlphaGo, AlphaCode, and AlphaProof. Reflection does not publish individual staff attributions on its site, so treat specific project credits as reported rather than company-confirmed.

• Misha Laskin — co-founder and CEO, formerly of Google DeepMind.
• Ioannis Antonoglou — co-founder, formerly of Google DeepMind, credited as an AlphaGo co-creator.
• Founding year — 2024, with reports placing the launch in March of that year.
• Team background — reported to include people who worked on Gemini, AlphaCode, and AlphaProof.
• Unconfirmed — the company does not publish a per-person project history.

How Much Has Reflection AI Raised, and from Whom?

Reflection AI confirmed on its own blog that it raised $2 billion to build frontier open intelligence. That announcement is the strongest primary source on the round.

The company names its investors directly: B Capital, Citi, CRV, Disruptive, DST, Eric Schmidt, Zoom Ventures, Lightspeed, Nvidia, Sequoia, 1789, "and others." Nvidia is the investor most often attached to the company in press coverage.

Valuation figures come from outside the company blog and should be read accordingly. Reflection's own news page links a CNBC segment dated April 23, 2026, in which the CEO confirms a round closed at a $25 billion pre-money valuation.

• Confirmed by Reflection — a $2 billion raise, announced on the company's own blog.
• Confirmed investor list (company-published) — B Capital, Citi, CRV, Disruptive, DST, Eric Schmidt, Zoom Ventures, Lightspeed, Nvidia, Sequoia, and 1789.
• Confirmed by the CEO on the record — a round closed at a $25 billion pre-money valuation (CNBC, April 23, 2026, linked from Reflection's own news page).
• Reported, not company-confirmed — an $8 billion valuation on the October 2025 round, and roughly $800 million of that round coming from Nvidia.
• Reported, not company-confirmed — an earlier $130 million total across a seed and Series A, disclosed when the company emerged from stealth in March 2025.

The Compute Deals Behind the Model

Reflection has signed unusually large compute agreements for a company whose first model, Beam, only shipped its weights on October 16, 2026. Its own news page links the coverage of each one.

The largest is with SpaceX. CNBC reported on June 22, 2026 that SpaceX signed a computing-power deal with Reflection worth up to $6.3 billion.

Bloomberg reported on July 14, 2026 that Nebius would sell $1 billion in AI capacity to Reflection. The Wall Street Journal reported in March 2026 that the company planned to spend billions on a data center in South Korea.

• SpaceX — a computing-power deal worth up to $6.3 billion (CNBC, June 22, 2026).
• Nebius — $1 billion in AI capacity (Bloomberg, July 14, 2026).
• South Korea — a reported multi-billion-dollar data center plan (Wall Street Journal, March 16, 2026).
• Public sector — Reflection was reported by Axios on May 22, 2026 to be powering the Genesis Mission.
• Why it matters to a buyer — compute commitments signal intent and scale. They are not evidence of model quality.

What Has Reflection AI Released So Far?

On October 5, 2026, Reflection AI announced Beam, its first open-weight model. Beam uses a sparse Mixture-of-Experts (MoE) architecture containing 501 billion total parameters and 23 billion active parameters, designed specifically for text-only coding, reasoning, and agentic workloads. Reflection pretrained the model on 23.8 trillion tokens using 6,144 NVIDIA GB300 GPUs, with up to 1M tokens of context after midtraining. For a detailed breakdown of benchmark data and architecture choices, read our full Reflection Beam explainer.

Reflection AI released Beam's model weights on Hugging Face on October 16, 2026, under an Apache 2.0 license in BF16, FP8, and NVFP4 formats. Developers can also access the Reflection Application Programming Interface (API) in public beta, which Reflection states is free to start. Reflection's documentation, checked October 16, 2026, lists no per-token rates for the API.

• Released: Asimov, Reflection AI's shipped product.
• Released: Beam, an open-weight 501-billion-parameter model (weights published October 16, 2026) supporting up to 1 million context tokens for reasoning and software engineering tasks.
• Not yet available: Beam through Microsoft Azure and Oracle, which Reflection's documentation lists as coming soon.
• Not found: Per-token API rates in Reflection's documentation, and independently verified benchmark scores.

When Beam's Open Weights Shipped

Reflection AI released the model weights for Beam on Hugging Face on October 16, 2026, under an Apache 2.0 license. The release followed an October 5 announcement. Reflection AI shipped the sparse Mixture-of-Experts (MoE) model, which uses 501 billion total parameters and 23 billion active parameters, in BF16, FP8, and NVFP4 formats alongside a speculative decoding drafter model.

The official developer documentation specifies a 1-million-token context window and a 128,000-token maximum output. Beyond Hugging Face, developers can query the Reflection Application Programming Interface (API) beta or deploy through partners like Amazon Web Services (AWS) SageMaker JumpStart. Learn more in our guide to Reflection Beam or our index of American open-weight AI models.

• Model name: Beam (501 billion total parameters, 23 billion active parameters)
• Announced: October 5, 2026
• Weights: Published October 16, 2026 on Hugging Face
• License: Apache 2.0
• API access: Public beta, free to start
• API price: Not listed in the documentation we checked

How Big Is Reflection AI?

Reflection does not publish a current headcount. The most reliable public figure is dated.

TechCrunch reported in October 2025 that the team numbered about 60 people, mostly AI researchers and engineers. That number is nearly a year old as of this writing.

Third-party company-data trackers publish 2026 estimates that disagree with each other by a wide margin. Those estimates are scraped, not disclosed, so this page does not repeat them as fact.

• Last credible reported figure — about 60 people (TechCrunch, October 2025).
• Current headcount — not published by the company.
• Third-party estimates — inconsistent across trackers; treat them as unconfirmed.
• What is clear — the company is hiring, and the compute commitments imply substantial growth in research staff.

What a Business Buyer Should Do About Reflection AI Now

You can test Beam through the Reflection AI public beta API or download its weights, but pricing and independent verification are still outstanding. Published benchmark numbers remain vendor-reported and unverified by third-party testing labs according to reporting from TechCrunch. At Layer3 Labs, we suggest piloting Beam on a bounded workload before moving any production pipeline onto it.

Treat Beam as an evaluation candidate rather than a committed infrastructure layer:

• Create an API key at the Reflection AI developer portal to test model behavior in the public beta.
• Read our Reflection Beam explainer to examine reported benchmark claims and architecture details.
• Check Reflection's serving guides before buying hardware: its single-server guide lists four NVIDIA B300 GPUs.
• Wait for independent testing groups to reproduce Reflection AI's scores on benchmarks such as SWE-bench Verified before swapping out existing coding assistants.
• Compare working alternatives alongside Beam in our guide to American open-weight AI models if you need self-hosted open models in production right now.

Status as of October 16, 2026

Reflection AI announced Beam on October 5, 2026, marking the debut of its first open-weight model. In Introducing Beam, the laboratory outlined a 501-billion-parameter Mixture-of-Experts (MoE) architecture designed for coding, reasoning, and agentic workflows. We updated this profile to incorporate Reflection AI's initial hardware specifications, cluster infrastructure, and self-reported performance figures, with detailed evaluation notes in our guide to Reflection Beam.

We updated this overview on October 16, 2026, after Reflection AI released the Beam weights under an Apache 2.0 license. We will add independent benchmark results and official token rates from the Reflection developer portal when they appear.