For most of 2026, the open-weight AI conversation has been dominated by Chinese labs — DeepSeek, Qwen, Z.ai's GLM series, Kimi. On October 5, a two-and-a-half-year-old Brooklyn startup most people have never heard of tried to change that. Reflection AI unveiled Beam, its first frontier open-weight model, and the pitch is specifically that it can match China's best open systems while using a fraction of the compute to run.
Quick summary: Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin (CEO) and Ioannis Antonoglou (president and CTO, a co-creator of AlphaGo), unveiled Beam on October 5, 2026 — a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens with a 1-million-token context window. Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning while using roughly one-quarter to one-third the inference compute. Full weights will be released under an Apache 2.0 license later this month. The company has raised about $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners at a recent $25 billion pre-money valuation.
Who is Reflection AI
Reflection was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both former staff research scientists at Google DeepMind — Laskin led reward modeling work for Gemini, and Antonoglou is credited as a co-creator of AlphaGo, the system that beat the world's top Go players back in 2016. Laskin is CEO; Antonoglou is president and chief technology officer. In roughly two years the company has raised about $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, most recently at a $25 billion pre-money valuation, and has reportedly locked in more than $7 billion in compute deals with SpaceX and Nebius for access to Nvidia's GB300 chips through 2029 (Sources: TechCrunch, Semafor).
What Beam actually is
Beam is a text-only mixture-of-experts model: 501 billion total parameters with only 23 billion active per forward pass, trained on 23.8 trillion tokens using what Reflection describes as high-compute reinforcement learning, with a 1-million-token context window. Reflection's own claim is that Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using roughly one-quarter to one-third as much inference compute, and that it beats Inkling — the model from Mira Murati's startup, released in July 2026 — on coding tests. The company says it's purpose-built for coding and agentic tasks, which is also where most of the open-weight competition from Chinese labs has concentrated this year.
An important caveat
Every one of those benchmark comparisons is Reflection's own claim, not yet independently verified by a third party — TechCrunch's own reporting explicitly flags that the performance claims "haven't been independently verified." It's also worth being precise about the release timeline: Beam was unveiled and benchmarked on October 5, but the full weights, technical report, and model card are scheduled to come out later this month, distributed through hyperscalers and neoclouds under an Apache 2.0 license. Until independent evaluators get their hands on the actual weights, treat the GLM-5.2 and Inkling comparisons as a company's marketing claim rather than a settled fact.
If you're choosing a model based on benchmark claims, wait for independent third-party evaluation rather than a lab's own numbers — this is true of every new model launch in 2026, not just Beam, and it's especially relevant here since the full weights aren't public yet.
Why Nvidia cares about a startup most people haven't heard of
Nvidia's interest in Reflection isn't incidental — Nvidia led a roughly $2 billion round into the company in October 2025 at an $8 billion valuation, and has backed the push toward a credible, Western, open-weight alternative to the Chinese open ecosystem ever since. That's a strategic interest beyond just returns: Nvidia sells the chips that run these models regardless of who builds them, but a world where the best open-weight models all come out of Chinese labs running on non-Nvidia or constrained hardware is a worse one for Nvidia's own compute business than a world where a company like Reflection, running on Nvidia GB300s, is competitive at the frontier of open models (Source: TechCrunch).
01What is Reflection AI's Beam model?
Beam is a 501-billion-parameter (23 billion active) open-weight mixture-of-experts language model from Reflection AI, unveiled October 5, 2026, with a 1-million-token context window. It's aimed at coding and agentic tasks and is positioned as a Western alternative to Chinese open models like DeepSeek, Qwen and GLM.
02Is Beam actually better than Chinese open models like GLM-5.2?
Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning while using about one-quarter to one-third the inference compute, but this is the company's own benchmark claim and has not yet been independently verified by third-party evaluators.
03When can I actually download Beam's weights?
Beam was unveiled and benchmarked on October 5, 2026, but Reflection says the full weights, technical report and model card will be released under an Apache 2.0 license later in October 2026, distributed through hyperscalers and neoclouds.
04Who founded Reflection AI?
Misha Laskin (CEO) and Ioannis Antonoglou (president and CTO), both former Google DeepMind staff research scientists. Antonoglou co-created AlphaGo; Laskin led reward modeling work for Gemini. They founded the company in March 2024.
Once Beam's weights are actually public and independently benchmarked, the fastest way to judge whether it lives up to Reflection's claims is to run it against the models you already use. LumiChats gives you access to 40+ leading models — proprietary and open-weight alike — in one place, so you don't have to take any single company's benchmark chart at face value.
