Reflection AI says its first model, Beam, has 501 billion parameters and uses only 23 billion of them for each token it writes. The Reflection AI Beam launch is a well-funded US lab’s debut, and its pitch is efficiency. We think that’s the right pitch for an open model. We also think nobody can check it yet, because the weights aren’t out. Wait for the download, then test 50 of your own items before you plan anything around it.
The numbers come from Reflection’s launch post, which describes Beam as a coding and agent model with open weights under Apache 2.0, due “later this month”. Every performance figure below is the vendor’s.
Key points
- What is Reflection AI Beam, and can anyone download it? No, not yet. Access today is a waitlist.
- Does a 3 to 4 times compute saving mean a 3 to 4 times smaller bill? No. Reflection says Beam scores comparably to Z.ai’s GLM-5.2 on advanced reasoning while using roughly 3 to 4 times less inference compute, and that the gap widens against models above 2 trillion parameters.
- Where does Beam sit against the leaders? Behind them, by Reflection’s own table. 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 that it lists.
- Who would actually want it? A firm that prefers a US-built model under a permissive licence, and doesn’t want Chinese-origin weights in its stack.
What is Reflection AI Beam, and can anyone download it?
No, not yet. Access today is a waitlist. The post says the model is in final red-teaming and evaluation, and that weights, a technical report and a model card follow later in October. There is no price and no hosted offering.
The design is a sparse mixture of experts, which is why 23 billion active parameters can sit inside a 501 billion total. It supports up to 1 million tokens of context. Apache 2.0 is a permissive licence, so a business can use and adapt the model commercially, subject to its terms. Read the licence text when it ships before telling anyone it’s free to deploy.
A buyer should care how many parameters fire per token. Think of a large firm where only the relevant specialists work on each job. The firm is big, but each job is cheap to staff. That’s the idea behind the design, and it’s why the compute claim is believable in principle. It’s also why the claim needs a price attached before it means anything to you. A 1 million token context is a spec on a page until someone shows it working on a long, messy file without losing the thread.
Two things the post leaves out matter here. It doesn’t give a release date beyond “later this month”. And it doesn’t say whether the training data and code will be open, only the weights. For a firm that cares about auditing what went into a model, that gap matters.
Does a 3 to 4 times compute saving mean a 3 to 4 times smaller bill?
No. Reflection says Beam scores comparably to Z.ai’s GLM-5.2 on advanced reasoning while using roughly 3 to 4 times less inference compute, and that the gap widens against models above 2 trillion parameters. It’s plausible. It’s also an estimate against a model its rivals have already moved past.
Read the footnote. Compute is estimated from active parameters and average output length. It leaves out prefill, attention and serving overhead, so it approximates work done and doesn’t measure cost. Your bill depends on how someone prices the hardware, and Reflection published no price.
Here’s a worked example with invented round numbers, ours and not a measurement. If a rival model cost you $1,000 a month to run for a task, a true 3x compute saving wouldn’t mean $333. Serving costs, idle capacity and engineering time don’t shrink with active parameters. It’s like swapping to a smaller engine and expecting the insurance, the garage and the driver to get cheaper too. The savings are real but smaller than the ratio suggests.
Where does Beam sit against the leaders?
Behind them, by Reflection’s own table. 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 that it lists. On Terminal-Bench v2.1 it gets 80.1, against 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash.
The gaps widen on harder tests. On DeepSWE v1.1, Beam scores 44.4 to Kimi K3’s 68.0 and DeepSeek V4.1 Flash’s 74.2. Reflection itself says Kimi K3 remains ahead on raw capability and that Beam’s edge is efficiency. That candour is useful, and it means the pitch of Beam as an answer to China’s open models depends on what you value.
For the leader in that table, see our Kimi K3 business test. For the comparison model, our GLM-5.2 benchmark test is a practical look at what running a model yourself involves. Open models are also showing up in security work, where GLM-5.2 cybersecurity results reportedly matched Mythos.
Who would actually want 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 more than a technology one, and for some buyers it’s decisive. Whether Beam is good enough to bother with, we can’t say from here.
The sceptic says a first model that trails the leaders and isn’t downloadable is a press release, not a product, and that a heavily funded lab can afford confident framing. That’s a fair read of the present state. Where we disagree is the efficiency angle. Most small firms will pay for hosting rather than build it themselves, and a model that does comparable work on fewer active parameters can be cheaper to host. We’d be wrong if independent tests showed Beam needing the same hardware as its rivals in practice, because then the claimed edge disappears.
Our guide to what AI costs Canadian businesses and our piece on cost per successful task show how to compare hosted and self-run options on total cost.
What should you do before planning around it?
Nothing yet, then something specific. Everything we know about Beam’s performance comes from Reflection’s own post. The weights aren’t public, so no outside party could reproduce a number, and the benchmark names are those the vendor chose. Several rows compare Beam to older rivals. The compute figure is a formula.
Once the weights land with the Apache 2.0 licence intact, pull 50 real items from your own work, the messy ones, and run them through Beam and whatever you use now. Count the ones your team accepts. Then ask a hosting provider for a price, because that’s the number a small firm will pay.
Our wait-and-see stance would flip fast if independent testers found Beam matching the leaders on messy, real coding tasks at a hosted price well below theirs. It would flip the other way if the weights slipped past October or the licence came with strings. Check back at the end of the month, and keep the 50 items. They’ll be useful the next time a lab announces a model with a nice efficiency 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?
No, not on Reflection’s own table. 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.
