Reflection Beam: 501B Open-Weight Model Aims at China's Best
Reflection Beam is Reflection AI's first open-weight model: 501 billion parameters, 23 billion active, Apache 2.0 weights promised later in October. For now it is a waitlist and a benchmark table.
- Beam is a sparse mixture-of-experts model with 501B total and 23B active parameters, pretrained on 23.8 trillion tokens, according to Reflection's announcement.
- Reflection claims reasoning scores on par with Z.ai's GLM-5.2 at 3 to 4 times less inference compute. TechCrunch notes these claims have not been independently verified.
- Weights, technical report and model card are promised "later this month." Today, only a select group of waitlist users can try the preview.
in this block
Reflection Beam is here, sort of. Reflection AI unveiled Beam on October 5, a 501 billion parameter open-weight model aimed at coding and agents, and says the weights land later this month under Apache 2.0. Right now it is a waitlist, a blog post and a very big benchmark table.
What actually happened
Reflection published a long post titled "Introducing Beam" on its official blog. The Brooklyn-based startup, founded in 2024 by two former Google DeepMind researchers, calls Beam a text-only "workhorse model" for enterprises, the public sector and developers.
TechCrunch confirmed the launch the same day and said it matched weekend reporting from Axios that a release was close. Per TechCrunch, citing PitchBook, Reflection has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed, with its last round at a $25 billion pre-money valuation.
The compute story is chunky. Reflection says pretraining ran in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs. The reinforcement learning run used 10.5K GB300 GPUs for four weeks and generated more than 100 million rollouts. TechCrunch adds that Reflection signed compute deals worth more than $7 billion with SpaceX and Nebius this summer.
The Reflection Beam scorecard
Every number below is self-reported by Reflection. Treat it like a project's own pitch deck: useful, but not gospel.
On coding, Beam posts 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1, versus 81.0 for GLM 5.2 on the latter. On reasoning it shows 97.8 on AIME 2026 and 90.5 on GPQA Diamond. On Humanity's Last Exam without tools it scores 36.2, behind GLM 5.2 at 40.5 and Kimi K3 at 46.9.
Reflection is upfront that Kimi K3 is still ahead on raw capability. Its argument is efficiency: comparable reasoning to GLM-5.2 while burning far fewer tokens and FLOPs. The post explains that its FLOP math is an estimate built from active parameters and generated tokens, not a measured serving cost.
Two practical details stand out. Reflection says midtraining extends Beam's effective context to 1 million tokens, and the model ships with a reasoning effort setting: lower effort for shorter answers, higher effort for hard tasks. That knob is how Reflection Beam is supposed to trade cost against quality.
TechCrunch frames Beam's closest US rival as Inkling from Mira Murati's Thinking Machines Lab. Beam beats Inkling on four coding tests where both report results, but Inkling is multimodal and Beam is text-only. Apples and slightly different apples.
Open weights "later this month" is a promise. Open weights on Hugging Face is a fact. Only one of those exists today.
Why degens and devs care
The open-weight race has been dominated by Chinese labs. We covered that wave in our Kimi K2.6 and Atria Dawn Preview pieces. A well-funded US lab shipping an Apache 2.0 model at this size changes the menu for anyone who wants to self-host an agent without a closed API.
Reflection also pitches "AI factories": institutions training Reflection models on their own data. TechCrunch says Reflection is testing a sovereign AI factory partnership with Shinsegae Group in South Korea. Expect hyperscaler and neocloud distribution partners at launch, per the company.
The safety section is notable too. Reflection says it trained a separate safety-and-alignment teacher and merged it into Beam via multi-teacher on-policy distillation. It promises to publish safety evals in the technical report and open-source its internal safety tests.
What to do as a reader
If you build agents, join the waitlist and wait for the weights before you rewrite your stack. The things that matter will be in the model card: license text, context limits in practice, hardware needs for 501B total parameters, and independent evals from third parties.
If you are a token degen, be careful. A hot AI launch reliably spawns copycat coins with the model's name. Reflection has not announced any token, and Reflection Beam is not a crypto project. Anything on a launchpad named after it is someone else's meme.
Bookmark the official blog and check back when the weights drop. Until then, Reflection Beam is a strong claim with a clear deadline, and "later this month" is exactly the kind of pending story we like to track. Not investment advice. Stay curious, ser.
Not financial advice. DYOR, ser.
