Raised $4.7 Billion. First Product, Two Years Later.
Beam is the West's most honest answer to China's open-weight wave: compete not on strength, but on cheapness. Every benchmark is self-reported and the efficiency comparison uses an estimation formula; the real bet is whether the 'intelligence density' road leads anywhere.
- 01THE FOUNDERS — In March 2024, two former DeepMind researchers founded Reflection AI in Brooklyn.
- 02THE MONEY — Roughly $4.7 billion raised, Nvidia alone contributing $800 million; the valuation multiplied roughly 45-fold in two years to $25 billion.
- 03THE LAUNCH — On October 5, Beam launched: a 501-billion-parameter MoE that concedes it trails Chinese models and competes on price.
WHAT HAPPENEDWhat Happened
On October 5, Reflection AI unveiled Beam: founded in Brooklyn in March 2024 by two former Google DeepMind researchers (CEO Misha Laskin, who led Gemini reward modelling, and CTO Ioannis Antonoglou, a co-creator of AlphaGo), the company released a public model for the first time in more than two years. The specs are solid: a sparse MoE architecture with 501 billion total parameters and only 23 billion active per token, pre-trained on 23.8 trillion tokens, a 1 million-token context window, text-only.
All comparison figures are self-reported and not independently verified: on hard reasoning, code and agent benchmarks it claims parity with Zhipu's GLM-5.2 (744 billion total parameters, 40 billion active) at one-third to one-quarter of the inference compute; its SWE-bench Verified score of 80.9 beats the Inkling model (77.6) in its own table. But the same materials concede that on most leaderboards it trails Kimi K3 and DeepSeek V4.1 Flash. The selling point is not peak performance; it is 'intelligence density' per unit of compute.
THE HONEST PLAYAdmitting it cannot win: the contrarian play
The usual script for a 2026 model launch is 'leads across the board' and 'new SOTA'. Beam's launch materials contain a different, startlingly honest sentence: on most leaderboards it trails Kimi K3 and DeepSeek V4.1 Flash. Writing that sentence into a press release takes courage — or calculation: when you know you cannot win on peak performance, you move the battlefield to value for money.
Its logic is 'intelligence density': the same intelligence for less compute. It claims parity with GLM-5.2 at one-third to one-quarter of the inference cost. But note: that efficiency comparison uses an estimation formula (2 × active parameters × average generated tokens), not measured bills. The cheapness is real, but the 'several times cheaper' figure itself was computed on an exam paper the company set itself.
THE EFFICIENCY ARITHMETICThe arithmetic behind 501 billion parameters
The training numbers are unusually concrete: 6,144 Nvidia GB300 NVL72 systems, pre-training completed in under four weeks at 92.3% goodput; plus more than 100 million RL rollouts across roughly 1 million synthetic code, agent and STEM environments, burning another 10,500 GB300s. Publishing a training ledger at this granularity is itself a signal: the bet is on engineering efficiency, not mystique.
But 'open weights' are still a futures contract: the weights are promised under Apache 2.0 within October; for now only a waitlist gets early access, citing 'final safety testing'. Released, open-sourced — and still unusable. The standard 2026 script.
THE $25B FIRST DELIVERYA $25 billion valuation hands in its first product
The money is more extravagant than the model: roughly $4.7 billion raised in total; an October 2025 $2 billion Series B at an $8 billion valuation, with Nvidia alone putting in $800 million; a 2026 round valuing it at about $25 billion — a roughly 45-fold increase in two years. The endorsement list includes Nvidia, Sequoia, Lightspeed, DST, Eric Schmidt and even Donald Trump Jr.'s 1789 Capital.
The compute contracts are harsher still: up to $6.3 billion to rent GB300s at SpaceX's Colossus 2 from July 2026, more than $1 billion committed to Nebius through 2029, and a 250MW sovereign AI plant with South Korea's Shinsegae Group. The WSJ calls it the 'DeepSeek of the West'. But DeepSeek was chased after it built something; Reflection was chased first, and delivered two years later.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISAdmit you cannot win first; then work out how to survive
The second half of the open-model race will not be about whose parameters are bigger, but whose ledger is thinner. Reflection's $25 billion valuation and $4.7 billion raise prove one thing: against Chinese models, the 'leads across the board' boast no longer flies — what flies is 'we are cheaper'. When a company dares to write 'we trail Kimi and DeepSeek' into its own press release, it is not conceding. It is changing tracks.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Official launch specs: 501 billion total parameters, 23 billion active, trained on 23.8 trillion tokens; claims parity with GLM-5.2 at one-third to one-quarter of the inference compute; weights promised under Apache 2.0 within October, currently waitlist-only.
PRIMARY SOURCESPrimary Sources
- The Decoder: Reflection's Beam becomes the most capable open-weight model built outside China (2026-10-05)
- Startup Fortune: Reflection AI launches Beam, an open model it says beats the West on efficiency
- How2Shout: Reflection AI Beam open-weight model specification roundup
- the-decoder.com: Original article
https://the-decoder.com/reflections-beam-becomes-the-most-capable-open-weight-model-built-outside-china/ - startupfortune.com: Original article
https://startupfortune.com/reflection-ai-launches-beam-an-open-model-it-says-beats-the-west-on-efficiency/
RELATEDRelated Intelligence
- Oklo Posts First Revenue of $1.86M as Q3 Loss Widens to $41.7M, Liquidity Hits $3.13B
- Moderna Swings to $162M Profit as Q3 Revenue Hits $1.31B, Crushing Estimates
- Monster Q3 Sales Hit a Record $2.56B on International Growth, but Margins Squeeze Shares
- BILL Q1 Profit Beats, but an Accounting Change Makes Revenue Look Like a Miss