Reflection AI

Reflection AI
Reflection AI, Inc.
Type	Private company
Industry	Artificial intelligence, Software development, Foundation models
Founded	March 2024
Founders	Misha Laskin (CEO), Ioannis Antonoglou (CTO)
Headquarters	New York City, New York, United States
Offices	New York, San Francisco, London, Washington, D.C.
Key people	Misha Laskin (CEO), Ioannis Antonoglou (CTO)
Products	Beam (open-weight MoE language model, 2026), Mirror CLI (open-source coding agent), Reflection API platform, Asimov (code research agent, 2025)
Valuation	$8 billion (October 2025); $25 billion pre-money for the round the CEO said had closed in April 2026
Total funding raised	More than $2.13 billion disclosed through October 2025; size of the 2026 round not disclosed
Employees	About 60 (October 2025); more than 100 AI researchers (per Lightspeed, August 2026)
Key partners	NVIDIA, SpaceX, Nebius, GMI Cloud, Shinsegae Group, Dell, U.S. Department of Energy
Website	reflection.ai

Reflection AI (which brands itself simply as Reflection) is an American artificial intelligence company that builds open-weight foundation models and the software around them. It was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, and it began as an autonomous coding company.^[1][2] Reflection emerged from stealth in March 2025 with $130 million in funding, then raised $2 billion at an $8 billion valuation in October 2025 in a round led by NVIDIA, when it repositioned itself as what TechCrunch called "America's open frontier AI lab" and a Western counterweight to Chinese open-weight developers such as DeepSeek.^[2][1][3] In April 2026 Laskin said on CNBC that the company's next round had closed at a $25 billion pre-money valuation.^[4]

On October 16, 2026, Reflection released the weights of its first open-weight model, Beam, under the Apache 2.0 license. Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning and agentic work.^[5][6] The company also sells access through its own API, distributes Beam through more than 40 partners, and has signed compute agreements with SpaceX and Nebius and public-sector agreements with the U.S. Department of Energy and the Department of Defense.^{[5][7][8][9][10]} Antonoglou was a co-author of AlphaGo, AlphaZero and MuZero at DeepMind, while Laskin led reward modeling for Gemini.^[1][11]

Reflection describes its mission as making "intelligence open and accessible to all." It positions itself as an open-weight alternative to closed frontier labs such as OpenAI and Anthropic, and as a counterweight to leading Chinese open-weight efforts such as DeepSeek and Qwen, pairing large-scale reinforcement learning with mixture-of-experts language model architectures.^[12][1][13]

What is Reflection AI?

Reflection's founding thesis was that the generality of large language models, combined with the capability gains of reinforcement learning, is the most promising path to autonomous agents. When it left stealth in 2025, its stated plan had two steps: first build a superintelligent autonomous coding system, then use that blueprint for other kinds of computer-based work.^[14][15] SiliconANGLE described the company as treating autonomous coding as a "root node": by mastering software, an AI system could bootstrap other forms of automation.^[3]

The company's focus widened over 2025 and 2026. Its first product, Asimov (July 2025), is a code-research agent for enterprise engineering teams. In October 2025 Reflection announced it had built a large-scale LLM and reinforcement-learning training platform capable of training mixture-of-experts models "at frontier scale" and would use it to release open models.^[16][13] By October 2026 its website described a full stack built around those models: open-weight models, open-source software (deployment containers, customization recipes and agent harnesses), serving through its API platform or in customers' own environments, and a forward-deployed solutions team.^[17]

Reflection's privacy policy gives its address as 61 9th Avenue in New York City.^[18] Its careers page lists positions in New York, San Francisco, London and Washington, D.C., and some job listings are based in Seoul.^[19] Lightspeed Venture Partners wrote in August 2026 that the company had assembled more than 100 AI researchers from DeepMind, OpenAI, Anthropic and Meta.^[20]

Who founded Reflection AI?

Reflection AI was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, who met inside Google DeepMind while working on Gemini: Antonoglou had a leadership role in Gemini post-training and Laskin was leading reward-model development. Laskin is chief executive officer and Antonoglou chief technology officer.^[1][15][14]

Misha Laskin

Michael "Misha" Laskin is the company's chief executive officer. On the Manifold podcast he described himself as Russian Israeli and said that after moving to the United States he became interested in physics and literature as a teenager; he majored in physics at Yale University with a double major in literature, and earned a PhD at the University of Chicago in theoretical many-body quantum physics.^[21] He then founded a Y Combinator-backed startup, which he said converged on building inventory-prediction systems for retailers and became a consulting business without a clear product.^[22][21] He has said that seeing AlphaGo changed his direction, after which he taught himself deep learning and reinforcement learning.^[21]

Laskin then joined the Berkeley AI Research lab at the University of California, Berkeley as a postdoctoral researcher with Pieter Abbeel, working on sample-efficient reinforcement learning. He is a co-author of CURL (Contrastive Unsupervised Representations for Reinforcement Learning) and the first author of Reinforcement Learning with Augmented Data (RAD), both published in 2020.^[23][24][22]

At Google DeepMind he worked on the Gemini project and led reward modeling, the part of reinforcement learning from human feedback that scores model outputs; Lightspeed specifies Gemini 1.5.^[1][20][22] He left DeepMind in 2024 to co-found Reflection.^[1]

Laskin has been the more public-facing co-founder. Around the October 2025 round he framed the open-weight strategy as a response to Chinese labs: "DeepSeek and Qwen and all these models are our wake-up call because if we don't do anything about it, then effectively, the global standard of intelligence will be built by someone else." He also said that "the most impactful thing is the model weights, because the model weights anyone can use and start tinkering with them," while the company would largely keep datasets and full training pipelines proprietary.^[1]

Ioannis Antonoglou

Ioannis Antonoglou is the company's chief technology officer. Born and raised in Thessaloniki, Greece, he holds an engineer's degree in electrical and computer engineering from the Aristotle University of Thessaloniki (2011) and a master's degree in artificial intelligence and machine learning from the University of Edinburgh (2012).^[25][11]

Antonoglou joined DeepMind in late 2012 as employee number 25 and the sixth member of the research team, before the company was acquired by Google. Over roughly twelve years there he co-authored a series of landmark systems:^[25][11]

• Deep Q-Networks (DQN), which learned to play Atari games from raw pixels (2013 and 2015 papers).
• AlphaGo, which defeated world champion Lee Sedol in 2016.
• AlphaGo Zero and AlphaZero, which learned Go, chess and shogi from self-play without human game data.
• MuZero, which learned to plan in Go, chess, shogi and Atari with a learned model of the environment, without being given the rules.
• Gemini, where he led the team responsible for RLHF, according to Endeavor Greece; Sequoia describes "a key leadership role in post-training efforts for Gemini."^[25][15]

In a March 2026 interview with the Turing Post, Antonoglou explained the shift toward building an open base model: "If you want to do reinforcement learning at the frontier, you need an extremely powerful base model to post-train." He said that "the whole Western ecosystem was missing a powerful open base model that we could even use to do reinforcement learning at scale."^[14]

What is the history of Reflection AI?

Founding (2024)

Laskin and Antonoglou left DeepMind and started working on Reflection in late February and early March 2024, according to the Turing Post; TechCrunch dates the launch to March 2024.^[14][1] The company spent its first year in stealth.^[14]

Stealth exit and early funding (March 2025)

Reflection AI emerged from stealth on March 7, 2025, saying it was focused on building autonomous coding systems. A Bloomberg report cited by PYMNTS said it had raised $130 million: a $25 million seed round led by Sequoia Capital and CRV, and a $105 million Series A led by Lightspeed Venture Partners and CRV.^[2] Other reported investors included Reid Hoffman, Alexandr Wang, SV Angel and NVIDIA's venture arm.^[26] TechCrunch and SiliconANGLE later gave the resulting valuation as $545 million; The SaaS News reported $555 million.^[1][3][26]

Sequoia described itself as the company's "first and long-term partner" and summarized the pitch as superintelligent agents that would do knowledge work, starting with autonomous coding agents that tackle well-scoped engineering tasks end to end.^[15]

Asimov launch (July 2025)

On July 16, 2025, Reflection AI introduced Asimov, a code-research agent; Sequoia called it the company's "first step" toward its long-term goal of superintelligence.^[27][16] Rather than generating code, Asimov answers questions about how a software system works, drawing on whole codebases and on non-code material such as architecture documents, GitHub threads and chat history.^[16] In blind tests with maintainers of large open-source projects, Reflection reported that Asimov's answers were preferred a majority of the time over Cursor Ask and Claude Code (Sonnet 3.7 and 4); one internal survey gave 82% for Asimov against 63% for Claude Code on Sonnet 4. The Decoder noted that this survey was internal and not independent.^[16][28][27]

$2 billion round led by NVIDIA (October 2025)

On October 9, 2025, Reflection AI announced it had raised $2 billion at an $8 billion valuation, about 15 times its valuation seven months earlier.^[1] SiliconANGLE reported that NVIDIA led the round.^[3] TechCrunch listed investors including NVIDIA, Disruptive, DST, 1789, B Capital, Lightspeed, GIC, Eric Yuan, Eric Schmidt, Citi, Sequoia and CRV; Reflection's own post also thanked Zoom Ventures.^[1][13] NVIDIA's contribution was later reported at roughly $800 million.^[29][30]

The round coincided with a public repositioning. In a post titled "Building Frontier Open Intelligence," Reflection said it had assembled a team whose members had worked on PaLM, Gemini, AlphaGo, AlphaCode and AlphaProof and contributed to ChatGPT and Character AI, and that it had "identified a scalable commercial model that aligns with our open intelligence strategy."^[13] Laskin told TechCrunch the team numbered about 60 people and that the company aimed to release a frontier language model trained on "tens of trillions of tokens" early in 2026, with revenue to come from large enterprises and from governments building sovereign AI systems.^[1]

GMI Cloud partnership (November 2025)

On November 20, 2025, GMI Cloud, an NVIDIA Reference Architecture Platform Cloud Partner, announced a collaboration under which Reflection would use GMI's U.S.-based GPU clusters to train its models. The two companies said they were also exploring AI factory and sovereign AI projects.^[31]

Investor talks (March 2026)

On March 2, 2026, the Financial Times reported that Reflection was courting investors at a valuation of more than $20 billion.^[32][33] On March 25, 2026, the Wall Street Journal reported that the company was in talks to raise $2.5 billion at a $25 billion valuation. According to that report, JPMorgan Chase was in talks to join through its Security and Resiliency Initiative, a program launched in December 2025 under which the bank said it would invest up to $10 billion in venture-backed startups, and existing investor Disruptive was expected to participate.^[29]

Shinsegae sovereign AI factory in South Korea (March 2026)

On March 16, 2026, Reflection AI and South Korea's Shinsegae Group announced a memorandum of understanding to build a sovereign AI cloud for Korean enterprises and agencies, centered on a 250-megawatt AI factory powered by Reflection's open-weight models and NVIDIA GPUs. Under the MOU Reflection would provide "chips, models, and full-stack engineering," while Shinsegae would handle the physical facility, real estate, power, permitting and financing. The companies said the project, to be executed in phases, had the support of the U.S. and Korean governments.^[34] Korea Economic Daily put the data center at $6.7 billion and said U.S. Commerce Secretary Howard Lutnick attended the signing with Laskin, Antonoglou and Shinsegae chairman Chung Yong-jin.^[35]

On April 17, 2026, Shinsegae said it had suspended partnership talks with OpenAI, announced about ten days earlier, to concentrate on its work with Reflection. That work includes applying AI to retail operations led by its E-Mart unit, from product sourcing and pricing to logistics, inventory and customer management.^[36]

Round closed at $25 billion pre-money (April 2026)

On April 23, 2026, Laskin confirmed on CNBC's "Squawk on the Street" that Reflection's latest funding round had closed at a $25 billion pre-money valuation.^[4][30] The amount raised in that round was not stated in the reports reviewed here.

Government agreements (May 2026)

On May 1, 2026, the U.S. Department of Defense said it had signed agreements with NVIDIA, Microsoft, Amazon Web Services and Reflection AI to deploy their AI technology and models on its classified networks, at Impact Level 6 and 7, for "lawful operational use."^[10] On May 22, 2026, Reflection said it had signed a memorandum of understanding with the Department of Energy to provide an open-source AI platform for the Genesis Mission and the department's 17 national laboratories.^[9] Reflection's website also says it is validated on the Dell AI Factory with NVIDIA for on-premises deployment.^[12]

Compute deals with SpaceX and Nebius (June and July 2026)

On June 22, 2026, CNBC reported that Reflection had signed an agreement with SpaceX for immediate access to NVIDIA GB300 chips at the Colossus 2 data center. According to materials viewed by CNBC, Reflection agreed to pay $150 million a month from July 1, 2026 through 2029, about $6.3 billion if the contract runs its full term; either party can end it on 90 days' notice after the first three months.^[7] On July 14, 2026, Reflection said Nebius had agreed to sell it more than $1 billion of computing capacity through 2029, also on NVIDIA GB300 chips; Nebius confirmed the deal.^[8][37] "The need for open models is clear, and this additional compute capacity will allow Reflection to continue to build and train frontier AI models at scale," Antonoglou said.^[37]

Beam (October 2026)

On October 5, 2026, Reflection announced Beam, "Reflection's first open-weight model," as a preview for a selected group of users while red-teaming and evaluations were finished; it updated the post's benchmark results on October 8.^[38] On October 16, 2026, it released the weights on Hugging Face under the Apache License 2.0.^[5][6] The release included BF16, FP8 and NVFP4 checkpoints, a DFlash2 speculative decoding drafter, the tokenizer and a reference implementation, and Mirror CLI, which Reflection calls "the open source harness we built and trained Beam on." Reflection said it was launching with "an ecosystem of 40+ partners."^[6][39] NVIDIA said Beam was trained on NVIDIA GB300 NVL72 systems.^[40] Details are in the Beam article.

Reported NVIDIA acquisition talks (October 2026)

On October 10, 2026, the Financial Times reported that NVIDIA was in early talks to acquire Reflection or to invest more in it. As summarized by SiliconANGLE, the options included an outright purchase, a larger stake possibly combined with more chips and computing power, or an acquihire in which NVIDIA would hire Reflection's staff and license its technology. SiliconANGLE said neither company had commented publicly.^[30]

How much funding has Reflection AI raised?

Across rounds whose amounts were disclosed, Reflection AI raised more than $2.13 billion through October 2025: $25 million in seed funding, a $105 million Series A and the $2 billion October 2025 round.^[2][1] Laskin said in April 2026 that a further round had closed at a $25 billion pre-money valuation; the reports reviewed here do not give its size.^[4]

What partnerships does Reflection AI have?

Laskin has tied this deal-making to a "sovereign AI" argument: governments and large enterprises want models they can own, run on their own infrastructure and customize, which an open-weight lab can supply.^[1][34]

What products does Reflection AI make?

Beam

Beam is Reflection's first open-weight model and, according to Reflection, the first in a series. It is a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, released under Apache 2.0.^[38][5] Reflection says it was pretrained on 23.8 trillion tokens on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs and then post-trained with a four-week reinforcement-learning run on 10,500 GB300 GPUs that generated more than 100 million rollouts.^[38] Its API identifier is Beam-1.0-501B-A23B, with a 1M-token context window and a knowledge cutoff of June 30, 2026.^[41] Reflection's benchmark figures are self-reported; for example, SWE-bench Verified is 80.9 in the launch post and 81.6% for the BF16 checkpoint on the model card.^[38][42] Reflection says Beam is competitive with larger open models such as GLM 5.2 while "frontier open models like Kimi K3 remain ahead on raw capability," and that Beam's advantage is efficiency at inference time.^[38]

Reflection API and Mirror CLI

Reflection offers Beam through a first-party API at platform.reflection.ai, opened as a public beta that is "free to start," with an OpenAI-compatible endpoint documented at developers.reflection.ai.^[5][41] Mirror CLI is Reflection's open-source terminal coding agent. Reflection says it built Mirror for internal use, with isolated workspaces, background sessions and Slack, Linear and GitHub integrations, used it to train Beam to work with coding harnesses, and then open-sourced it.^[5]

Asimov

Asimov is Reflection's code-research agent, launched in July 2025 and described by Antonoglou as "something like Deep Research, but it's for your engineering systems."^[27] Where most coding tools focus on code generation in an editor, Asimov is designed to answer questions about how a software system works and why it was built a particular way.^[16]

Architecture

Asimov uses a multi-agent architecture with two roles:^[16][28]

• Retrievers: many small long-context agents that pull relevant information from a large codebase and its surrounding material.
• Combiner: one larger short-context reasoning agent that synthesizes the retrievers' findings into an answer.

Sequoia wrote at launch that the release was "powered by third-party models" while Reflection trained its own; The Decoder reported that Asimov ran on open-source models that Reflection had further trained with reinforcement learning on human-annotated and synthetic data.^[16][28]

Capabilities

1. Single source of truth for engineering knowledge. Asimov ingests entire codebases, architecture documents, GitHub threads and chat history and builds a persistent memory of how a system works and why decisions were made.^[16]
2. Team-wide editable memory. Engineers can add to Asimov's knowledge with commands such as "@asimov remember X works in Y way," and role-based access controls decide who can edit that memory.^[16][28]
3. Deep-research-style answers. Sequoia summarized the product as "Deep Research for code comprehension."^[16]

Deployment and access

Asimov is deployed inside customers' own virtual private clouds, and Reflection says customer data is not used for training.^[28][27] In March 2026 the Turing Post reported that onboarding still routed users to a waitlist that could not actually be joined.^[14]

Earlier autonomous coding agents

When Reflection left stealth, Sequoia described its product as autonomous coding agents that integrate with an organization's codebase and workflows and take on well-scoped tasks such as refactoring, debugging, documentation and infrastructure work, reading, writing, testing and deploying code.^[15]

Is Reflection AI open source?

Reflection describes itself as an open-weight lab. Its website lists three commitments: releasing open weights for its models, publishing the research behind them as papers and technical reports, and open-sourcing software, including reinforcement-learning tools and environments.^[12] In October 2025 Laskin said the company would largely keep datasets and full training pipelines proprietary.^[1]

Until October 2026 the commitment was untested. In March 2026 the Turing Post noted that Reflection's verified Hugging Face organization showed no public models or datasets and that its blog had two posts.^[14] Beam's release on October 16, 2026 put open weights, a tokenizer, a reference implementation and Mirror CLI under Apache 2.0, alongside the Beam Safety Technical Report.^[5][43] At release the Hugging Face organization listed four Beam repositories and no datasets.^[39]

What are the key milestones in Reflection AI's timeline?

What is Reflection AI's technical approach?

Reinforcement learning on language models

Reflection's founders come from DeepMind's reinforcement-learning lineage, from DQN through MuZero, and from Gemini post-training.^[11][15] Laskin has said that reinforcement learning needs a task distribution to learn from and a reliable way to verify that tasks were solved, and that after Gemini 1.5 launched he thought it was time to scale reinforcement learning on top of language models.^[21] Lightspeed described Reflection's early coding systems as evaluating multiple possible solutions, learning from failures and refining their approach, rather than only predicting the next line of code.^[20]

Antonoglou told the Turing Post that frontier reinforcement learning requires a strong base model to post-train, which is why the company decided to pretrain its own open model.^[14] For Beam, Reflection reports a reinforcement-learning run with a maximum rollout context of 256K tokens, about 1.3 billion sandboxes for training and grading, and nearly one million coding, agentic and STEM environments.^[38]

Mixture-of-experts at frontier scale

Reflection's training stack is built around mixture-of-experts architectures. In October 2025 it said it had built a platform capable of training "massive Mixture-of-Experts (MoEs) models at frontier scale," and Laskin said the first model would be trained on "tens of trillions of tokens."^[13][1] Beam, at 501 billion total and 23 billion active parameters pretrained on 23.8 trillion tokens, was the first public result.^[38]

Safety and openness

Reflection argues that openness and safety reinforce each other. In "Building Frontier Open Intelligence" it wrote that "the answer to AI safety is not 'security through obscurity' but rigorous science conducted in the open," and said it was investing in pre-release capability and risk evaluations, security research and responsible deployment standards.^[13] For Beam it published a safety technical report under what it calls its Open Safety Framework, rating Beam "Limited" in all four systemic-risk categories it assesses (biological weapons, chemical weapons, loss of control and cyber offense), meaning in Reflection's definition no material shift in the offense-defense balance relative to comparable open-weight models. These are Reflection's own determinations.^[43]

Who works at Reflection AI?

Reflection grew from its two founders in 2024 to about 60 people by October 2025, mostly researchers and engineers in infrastructure, data training and algorithm development.^[1] Lightspeed wrote in August 2026 that the company had more than 100 AI researchers from DeepMind, OpenAI, Anthropic and Meta, with more than 300,000 combined citations.^[20] Reflection says its team includes people who worked on PaLM, Gemini, AlphaGo, AlphaCode and AlphaProof and contributed to ChatGPT and Character AI.^[13]

Leadership

Who are Reflection AI's competitors?

Reflection AI has competed in two adjacent markets: coding agents and open-weight foundation models.

Autonomous coding

In coding, the Turing Post pointed to Anthropic's Claude Code, OpenAI's Codex and Cognition AI's Devin as faster-shipping competitors.^[14] Sequoia's Asimov launch post compared it with Cursor Ask, from Cursor maker Anysphere, and with Claude Code; SiliconANGLE noted that Reflection's comparison did not include OpenAI's models, GitHub Copilot or Devin.^[16][27]

How does Asimov compare with Devin and Cursor?

Most public benchmarks of coding agents, such as SWE-bench Verified, measure end-to-end code generation rather than code comprehension, so Asimov is usually compared with other tools qualitatively:

Frontier and open-weight foundation models

Since its October 2025 repositioning, Reflection has compared itself with both closed labs and open-weight developers. The Turing Post named DeepSeek, Qwen, Mistral AI, the Allen Institute for AI's OLMo, OpenAI's gpt-oss and NVIDIA's Nemotron family as the main open-model contenders, alongside the closed labs.^[14] In its Beam launch post Reflection compared Beam with Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash.^[38]

How has Reflection AI been received?

The October 2025 repositioning drew support from American technology figures. David Sacks, the White House AI and crypto czar, wrote: "It's great to see more American open source AI models." Clem Delangue, chief executive of Hugging Face, called the round "great news for American open-source AI" but added that "the challenge will be to show high velocity of sharing of open AI models and datasets."^[1]

Before Beam shipped, commentators questioned whether the company's output matched its valuation. In March 2026 the Turing Post wrote that Reflection had not released a public frontier model, that Asimov remained behind a waitlist and that its Hugging Face organization was empty, while acknowledging the strength of its founders, team and capital.^[14] AInvest described the gap between the reported $20 billion-plus valuation and the absence of a public model as a "glaring credibility gap."^[33] Later in 2026, CNBC noted that the company had built momentum with government and national-security customers before releasing a public frontier model.^[7]

Asimov drew privacy questions. MIT professor Daniel Jackson told Wired that the idea was promising but unproven and that it would be "reading all these private messages" between developers, as reported by SiliconANGLE.^[27]

Korea Economic Daily described the Shinsegae project as part of the Trump administration's policy of exporting the U.S. AI technology stack.^[35]

See also

• Beam (Reflection)
• AI agent
• Reinforcement learning
• Reinforcement learning from human feedback
• Large language model
• Mixture of experts
• Open weights
• AlphaGo
• AlphaZero
• Google DeepMind
• Cognition AI
• Cursor
• Magic.dev
• DeepSeek
• Anthropic
• OpenAI
• NVIDIA
• Artificial general intelligence

Notes

• This article summarizes publicly reported information about Reflection AI as of October 17, 2026. Funding talks and acquisition talks that have not been confirmed by the company are described as reported.

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21Rebecca Bellan. "Reflection AI raises $2B to be America's open frontier AI lab, challenging DeepSeek". TechCrunch. October 9, 2025.
2. ^1 ^2 ^3 ^4 ^5 ^6PYMNTS. "Reflection AI Raises $130 Million to Build Autonomous Coding Systems". March 7, 2025.
3. ^1 ^2 ^3 ^4 ^5 ^6SiliconANGLE. "Reflection AI lands $2B at $8B valuation to expand frontier AI infrastructure and safety research". October 9, 2025.
4. ^1 ^2 ^3 ^4CNBC. "Reflection CEO confirms latest funding round closed at $25 billion pre-money valuation". Squawk on the Street (video). April 23, 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Reflection. "Beam is now available". October 16, 2026.
6. ^1 ^2 ^3Reflection (@reflection_ai). Post announcing the open-weight release of Beam. X. October 16, 2026.
7. ^1 ^2 ^3 ^4Deirdre Bosa. "SpaceX signs computing power deal with open-source AI startup Reflection worth up to $6.3 billion". CNBC. June 22, 2026.
8. ^1 ^2 ^3Dina Bass. "Nebius to Sell $1 Billion in AI Capacity to Reflection". Bloomberg News (via Bloomberg Law). July 14, 2026.
9. ^1 ^2 ^3Weslan Hansen. "DOE Taps Reflection AI for Genesis Mission". MeriTalk. May 22, 2026.
10. ^1 ^2 ^3Ram Iyer. "Pentagon inks deals with Nvidia, Microsoft, and AWS to deploy AI on classified networks". TechCrunch. May 1, 2026.
11. ^1 ^2 ^3 ^4Chess Programming Wiki. "Ioannis Antonoglou". Accessed October 17, 2026.
12. ^1 ^2 ^3 ^4Reflection. "About". Accessed October 17, 2026.
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7Reflection. "Building Frontier Open Intelligence". Company blog. Accessed October 17, 2026.
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Turing Post. "Reflection AI: $25B Valuation, SpaceX Deal & the Open-Weight Bet". Original analysis March 2026, with a later update. Accessed October 17, 2026.
15. ^1 ^2 ^3 ^4 ^5 ^6Sequoia Capital. "Partnering with Reflection: Toward Superintelligence, with Autonomous Coding". March 2025.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Sequoia Capital. "Reflection AI Launches Asimov: Breakthrough Agent for Code Comprehension". July 2025.
17. ^Reflection. Official website. Accessed October 17, 2026.
18. ^Reflection. "Privacy Policy". Accessed October 17, 2026.
19. ^Reflection. "Careers". Accessed October 17, 2026.
20. ^1 ^2 ^3 ^4Lightspeed Venture Partners. "Open Intelligence for Everyone: Our Investment in Reflection AI". Originally published March 2025, updated August 2026.
21. ^1 ^2 ^3 ^4Steve Hsu. "Transcript: Misha Laskin, Reflection.ai: From Physics to SuperIntelligence". Manifold podcast. March 13, 2025.
22. ^1 ^2 ^3Misha Laskin. Personal website. Accessed October 17, 2026.
23. ^Aravind Srinivas, Michael Laskin, Pieter Abbeel. "CURL: Contrastive Unsupervised Representations for Reinforcement Learning". arXiv:2004.04136. 2020.
24. ^Michael Laskin et al. "Reinforcement Learning with Augmented Data". arXiv:2004.14990. 2020.
25. ^1 ^2 ^3Endeavor Greece. "From AlphaGo to Autonomous Agents: The Journey of Ioannis Antonoglou". Accessed October 17, 2026.
26. ^1 ^2 ^3The SaaS News. "Reflection AI Raises $130 Million in Funding". March 10, 2025.
27. ^1 ^2 ^3 ^4 ^5 ^6SiliconANGLE. "Reflection AI's autonomous coding agent Asimov learns from more than just code". July 16, 2025.
28. ^1 ^2 ^3 ^4 ^5 ^6 ^7The Decoder. "Reflection unveils Asimov: an AI agent built to track every step of software development". July 2025.
29. ^1 ^2 ^3 ^4Tech Startups. "Nvidia-backed startup Reflection eyes $2.5B round at $25B valuation as U.S. open-source AI push takes on China". March 26, 2026 (citing The Wall Street Journal).
30. ^1 ^2 ^3SiliconANGLE. "Nvidia reportedly in early talks to buy, acquihire or invest more in Reflection AI". October 11, 2026.
31. ^1 ^2GMI Cloud. "Reflection AI and GMI Cloud Partner to Accelerate the Future of AI-Driven Software Engineering". PR Newswire. November 20, 2025.
32. ^1 ^2Reflection. "Newsroom". Accessed October 17, 2026. Lists the Financial Times article "Nvidia-backed 'open' AI start-up courts investors at $20bn-plus valuation" (March 2, 2026) and Dell's release "Dell Technologies Closes the Gap Between AI Ambition and AI Outcomes" (May 18, 2026).
33. ^1 ^2AInvest. "Reflection AI Faces Critical March 2026 Test as $20 Billion Valuation Hinges on Public Model Release". March 2026.
34. ^1 ^2 ^3Reflection AI and Shinsegae Group. "Reflection and Shinsegae Group to Build Korean Sovereign AI Factory". PR Newswire. March 16, 2026.
35. ^1 ^2KED Global (Korea Economic Daily). "Shinsegae, US startup Reflection AI agree to build $6.7 bn data center in South Korea". March 17, 2026.
36. ^1 ^2Seoul Economic Daily. "Shinsegae Halts OpenAI Talks, Focuses on Reflection AI Partnership". April 17, 2026.
37. ^1 ^2The Next Web. "Reflection signs a $1bn AI compute deal with Nebius, its second big capacity grab in a month". July 2026.
38. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Reflection. "Introducing Beam: Reflection's 501B open-weight model". October 5, 2026 (results updated October 8, 2026).
39. ^1 ^2Reflection. "reflection" organization. Hugging Face. Accessed October 17, 2026.
40. ^1 ^2NVIDIA AI (@NVIDIAAI). Post on Beam's launch. X. October 16, 2026.
41. ^1 ^2Reflection Developer Docs. "Models". Accessed October 17, 2026.
42. ^Reflection. "Beam-1.0-501B-A23B" model card. Hugging Face. Accessed October 17, 2026.
43. ^1 ^2Reflection. "Beam Safety Technical Report". Accessed October 17, 2026.
44. ^1 ^2Awesome Agents. "SWE-Bench Coding Agent Leaderboard 2026". Accessed October 17, 2026.

External links

• Official website
• Reflection on Hugging Face
• Reflection developer documentation
• Reflection AI careers and Reflection Jobs on Ashby
• Reflection AI on LinkedIn
• Reflection AI on X (Twitter)