Reflection says its first open-weight model matches China’s GLM-5.2 with three to four times less compute. Nobody outside can test the claim yet.
Nvidia-backed Reflection AI unveiled Beam, its first open-weight model. The US startup built the model to compete with Chinese open models in coding and agentic tasks. Beam has 501 billion total parameters and activates 23 billion per task. Reflection says it’s competitive with Z.ai‘s GLM-5.2 and approaching Alibaba‘s Qwen 3.8-Max.
Weights, a technical report, and a model card arrive later this month under an Apache 2.0 license. Until then, a select group has early access, and nobody outside Reflection can verify the benchmark numbers.
What’s Happening & Why It Matters
What Reflection Claims
Reflection says Beam scores close to GLM-5.2 on advanced reasoning benchmarks with three to four times less inference compute. GLM-5.2 has about 744 billion total parameters and activates 40 billion. Against models above two trillion parameters, such as Qwen 3.8-Max, the company says the efficiency gap grows.
Reflection concedes limits. Moonshot’s Kimi K3 still leads on raw capability, and Beam trails the best closed models from OpenAI and Anthropic, CEO Misha Laskin said. The pitch is efficiency: strong coding and agent performance at a lower running cost. Reflection says it trained Beam end to end from scratch.
A Pitch to Governments and Companies
Laskin pitches Beam to companies and governments that want to control where AI runs. Today’s strongest open models come from Chinese labs. Axios reports open-weight models, most of them Chinese, have taken Silicon Valley by storm. An American model running on Nvidia GPUs, Axios says, gives buyers a cheaper alternative to Anthropic, OpenAI, and Google.
Reflection’s page says it is a consortium member of the Department of Energy‘s Genesis Mission, serving the 17 national laboratories. Axios’s sources say other Western open-weight models arrive this month. TF covered Beijing’s side of the contest in Xi Pitches Open-Source AI Alliance After Amodei Warning, where Xi pitched free access to Chinese open models for BRICS members.
The Money Behind Beam
Laskin and Ioannis Antonoglou, both former Google DeepMind researchers, founded Reflection in March 2024. In October 2025, it raised a $2 billion (€1.8 billion) Series B at an $8 billion (€7 billion) valuation. Nvidia led $800 million (€700 million) of the round, with Eric Schmidt, Citigroup, 1789 Capital, Lightspeed, and Sequoia Capital joining.
Startup Fortune, citing Axios, puts the valuation near $25 billion (€22 billion) today. It also says Reflection has more than $7 billion (€6.2 billion) in compute committed from SpaceX and Nebius. Nvidia’s stake fits a pattern. TF covered the chipmaker’s open-ecosystem bets in Nvidia Buys Hugging Face for $12.9 Billion, and open models run on its GPUs.
The Safety Question
Axios notes open models can be harder to monitor and regulate than closed frontier systems. Once weights are public, anyone can download and modify them, with no corporate safety layer attached. TF reported the risk in A Chinese AI Model Gave Bioweapon Instructions After a Jailbreak, where researchers bypassed safety controls on two Moonshot models.
Reflection says Beam’s weights ship after a final red-team review. Coverage doesn’t describe the review’s scope. Apache 2.0 allows commercial use and modification, so enterprises can adopt and modify the model.
What Happens at Release
Reflection plans to release the weights, a technical report, a model card, and developer materials later this month. Cloud distribution partners and open-source library integrations arrive at launch. Independent benchmark runs will follow within days and test whether Beam’s efficiency holds outside Reflection’s chart. Until then, the claim is a benchmark table and a $25 billion bet.
TF Summary: What’s Next
Reflection will release Beam’s weights under Apache 2.0 later in October, with no exact date named. Independent benchmark results come after release. Axios’s sources expect more Western open-weight models this month. Kimi K3 stays ahead on raw capability, by Reflection’s account.
MY FORECAST: Expect independent testers to confirm much of the efficiency claim and rank Beam below Kimi K3 on raw capability, matching Reflection’s concession. Expect at least one more Western open-weight release this month, per Axios’s sources. US agencies and contractors will cite Beam as a domestic alternative to Chinese models. Watch the license and red-team details at release, because they will decide whether enterprises adopt it.
Related Stories
- Nvidia Buys Hugging Face for $12.9 Billion
- A Chinese AI Model Gave Bioweapon Instructions After a Jailbreak
- Xi Pitches Open-Source AI Alliance After Amodei Warning
- Beam claims: Every performance and efficiency figure comes from Reflection’s announcement as relayed by outlets, and none is independently tested. The $25 billion valuation and $7 billion compute commitments come from Axios via Startup Fortune, not from the company.