Reflection Beam: what it is and when you can use it
2026-10-06 · 4 min read
Reflection Beam is a new open-weight AI model from Reflection, a US lab started in 2024 by two former Google DeepMind researchers. You can't download it yet: the weights, model card and technical report are due "later this month" under the Apache 2.0 license, and until then the only way in is a waitlist at platform.reflection.ai.
The announcement
On Monday, @reflection_ai introduced Beam as "a highly efficient agentic open model with 501B total parameters and 23B active." The post says it was trained from scratch and "advances the Western open frontier on coding & agentic tasks." It passed 1.4 million views within a day.
Co-founder @real_ioannis (Ioannis Antonoglou) added the backstory in a longer post. Pretraining started in July, the company has grown tenfold in a year, and he says the team is "already training the next, larger model."
The most useful outside reaction came from @ArtificialAnlys, the group that runs independent model benchmarks. It says Reflection gave it early access and it is testing Beam now. Its early read is that Beam looks like "one of the most token-efficient open models we've seen for its level of intelligence." That isn't a verdict, but so far it's the only signal that doesn't come from Reflection.
What Beam is, in plain terms
Beam is a mixture-of-experts model. It holds 501 billion parameters, but only 23 billion switch on for any given word it writes. That gives you a lot of stored knowledge while keeping each answer relatively cheap to compute. Other specs from Reflection's launch post:
- Text only, no images or audio
- A context window of 1 million tokens
- Pretrained on 23.8 trillion tokens, then trained with over 100 million reinforcement learning attempts on coding, reasoning, tool use and web search
- Apache 2.0, so businesses can use it commercially and fine-tune it
The benchmarks, read honestly
Reflection's own table is worth reading before the headlines. In every row of the comparison it published, Beam scores below the model it sits next to:
| Test | Beam | Compared model |
|---|---|---|
| SWE Bench Pro v2-Hard | 77.2 | GLM 5.3: 84.3 |
| Terminal Bench v2.1 | 80.1 | DeepSeek V4.1 Flash: 90.6 |
| AIME 2026 | 97.8 | GLM 5.2: 99.2 |
| GPQA Diamond | 90.5 | Kimi K3: 93.5 |
Reflection published those numbers itself, so it isn't hiding anything. The pitch is that Beam gets close to the leaders while doing far less work per answer. Its headline claim is that Beam scores about the same as Z.ai's GLM-5.2 with three to four times less inference compute. The Next Web pointed out that this figure is an approximation rather than a measurement, and that nobody outside Reflection has checked any of the scores yet. So the Artificial Analysis results will matter a lot, and they aren't out.
The other angle is geography. The best open models right now mostly come from Chinese labs: GLM, Kimi, DeepSeek, Qwen. Reflection, backed by Nvidia and valued at $25 billion before this round, is pitching Beam as the American answer, aimed at companies and governments that want a model they can run themselves.
Who can use it, and when
Right now, developers and companies on the waitlist. Once the weights land this month, anyone can download them. At 501 billion parameters, though, this won't run on an office PC or a single graphics card. Most people will reach Beam through cloud providers that host it, and Reflection says hyperscalers and smaller GPU clouds will carry it at launch. No per-token price has been announced.
What it means for a small business
Most owners will never download Beam, and they don't need to for the trend to reach them. Every efficient open model puts pressure on what the closed models charge for routine work like sorting emails, drafting quotes, or summarizing calls. When "good enough" gets three or four times cheaper to run, the cost of automating dull back-office tasks keeps falling, whether you end up on Beam or not.
For this month, I'd wait for the independent numbers and not switch anything because of a launch table. A better use of the time is figuring out which of your tasks need a top model and which a cheaper one could handle. New Face Design's free process audit sorts that out for Fox Valley businesses, task by task, before you pay for any model.