Say a rival model costs you $1,000 a month for a task. Reflection AI Beam claims roughly 3 to 4 times less inference compute for comparable reasoning, so does your bill fall to $333? It doesn’t, and nobody can test the claim yet anyway: Beam’s weights aren’t out. Reflection, a well-funded US lab, says its first model has 501 billion parameters and uses only 23 billion per token. Efficiency is the right pitch for an open model. Wait for the download, then test 50 of your own items before planning around it.
Key points
- Beam is an open-weight coding and agent model under Apache 2.0, due “later this month”. Access today is a waitlist.
- The compute figure is a formula and leaves out serving overhead. No price exists.
- Reflection’s own table puts Beam behind Kimi K3 and DeepSeek V4.1 Flash.
- Its clearest buyer prefers US-built weights under a permissive licence.
Why 3x less compute isn’t a third of the bill
Reflection’s launch post says Beam scores comparably to Z.ai’s GLM-5.2 on advanced reasoning at roughly 3 to 4 times less inference compute, with the gap widening against models above 2 trillion parameters. The footnote matters: compute is estimated from active parameters and average output length, leaving out prefill, attention and serving overhead. It approximates work done. It doesn’t measure cost.
In our invented example, serving costs, idle capacity and engineering time don’t shrink with active parameters, so the $1,000 lands well above $333. Fit a smaller engine and the insurance, the garage and the driver cost the same. The saving is real and smaller than the ratio. And GLM-5.2 is a model rivals have already moved past.

What Reflection AI Beam actually is today
A waitlist. The model is in final red-teaming and evaluation, with weights, a technical report and a model card promised later in October. No price, no hosted offering. It’s a sparse mixture of experts, which is how 23 billion active parameters sit inside 501 billion total: a big firm where only the relevant specialists work each job. It supports up to 1 million tokens of context, a spec on a page until someone shows it holding the thread on a long, messy file.
Apache 2.0 permits commercial use and adaptation, subject to its terms. Read the licence text when it ships. The post gives no firm date beyond “later this month” and doesn’t say whether training data and code will be open. For firms that audit what went into a model, that gap counts.
The vendor’s own scoreboard
Every figure here is Reflection’s. Beam scores 80.9 on SWE-Bench Verified and 65.5 on SWE-Bench Pro v1, ahead of the US open models Inkling and Nemotron 3 Ultra. On Terminal-Bench v2.1 it gets 80.1, against 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash. On DeepSWE v1.1 the gap widens: 44.4 to Kimi K3’s 68.0 and DeepSeek’s 74.2. Reflection itself says Kimi K3 stays ahead on raw capability and Beam’s edge is efficiency. That candour helps, and it means Beam as an answer to China’s open models depends on what you value.
See our Kimi K3 business test for the leader, and our GLM-5.2 benchmark test for what running a model yourself involves. Open models are showing up in security work too, where GLM-5.2 cybersecurity results reportedly matched Mythos.

Who should keep an eye on it
A firm that prefers a US-built model under a permissive licence and doesn’t want Chinese-origin weights in its stack. That’s a procurement preference, and for some buyers it’s decisive. The sceptic calls a trailing, undownloadable first model a press release dressed as a product. Fair, for now. We part ways on efficiency: most small firms pay for hosting, and fewer active parameters for comparable work can mean cheaper hosting. If independent tests show Beam needing the same hardware as rivals, that edge disappears. What AI costs Canadian businesses and cost per successful task show how to compare hosted and self-run options.
Your 50-item plan for Reflection AI Beam
Do nothing yet. When the weights land with Apache 2.0 intact, pull 50 real items, the messy ones, and run them through Beam and your current model. Count what your team accepts, then get a hosting quote, because that’s the number a small firm pays. We’d turn positive fast if independent testers found Beam matching the leaders on real coding tasks at a clearly lower hosted price, and negative if the weights slip past October or the licence comes with strings. Keep the 50 items for the next lab with a nice ratio and no price.
Frequently asked questions
Is Reflection AI’s Beam open source?
Reflection says it will release Beam’s weights under the Apache 2.0 licence later this month. That makes it open-weight. Whether the training data and code are open is not stated in the post. A European option is also on the table, and Mistral Large 4 is open-weight.
Can I use Beam today?
Only through a waitlist. Reflection says the model is in final red-teaming and evaluation, with weights, a model card and a technical report to follow.
Is Beam better than Chinese open models?
Reflection’s own table says it trails them. It leads some US open models and is comparable to GLM-5.2, but Kimi K3 and others score higher on several tests. Its pitch is lower compute for similar reasoning.
Archive entry dated October 6, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.