Oct 6, 2026 · 12 min · 8 segments
Tuesday morning AI and cyber briefing. Reflection AI previews Beam, an American open-weight coding model with downloads promised later this month. OpenAI starts adding its invisible textGrain…
Reflection AI is a young American company valued at around twenty-five billion dollars that has set out to build open models in the United States.
On Monday, it announced Beam, its first open weight model.
Open weights means the company will publish the trained model itself so anyone can download a copy, run it on their own machines, and change it.
Here's the catch, though.
Right now, Beam is only available to a small group through a waiting list.
It's still going through final safety testing, and the downloadable files, the technical write-up, and the tools for tuning it are promised for later this month.
It'll come under Apache two, a widely used license that allows commercial use without asking permission.
Reflection built it mainly for writing code and for agent work, meaning tasks where the AI takes a string of actions by itself rather than just answering a question.
Its own results put it close to last generation's leading Chinese open model while using about a quarter to a third of the computing power to get there and still behind the strongest open models available today.
Those are the company's own numbers, and nobody outside has checked them.
I'd like Tim's view on whether this is something to act on or just a date for the diary.
And for now, it's a promise because the files aren't out and the scores are self-reported.
So I'd treat it as a date for the diary, not a buying decision.
Where it really matters is for organizations whose rules say no Chinese-made models because until now, most of the strongest open options came from China.
If that's you, get a test ready on your own coding jobs so you can run it the week the files land and compare it against whatever you use today.
Isn't the American label a bit of a marketing line, though? A model doesn't care where it was trained.
The model doesn't care, but plenty of procurement and legal teams do.
And for them, the country of origin is a written rule rather than a matter of taste.
That's the real market for Beam.
If your policy has no such rule, judge it purely on how it does with your own work once it's out and ignore the flag on the box.
So a diary note for most people and a proper test plan if origin is on your rule book.
Reflection AI is a young American company valued at around twenty-five billion dollars that has set out to build open models in the United States.
On Monday, it announced Beam, its first open weight model.
Open weights means the company will publish the trained model itself so anyone can download a copy, run it on their own machines, and change it.
Here's the catch, though.
Right now, Beam is only available to a small group through a waiting list.
It's still going through final safety testing, and the downloadable files, the technical write-up, and the tools for tuning it are promised for later this month.
It'll come under Apache two, a widely used license that allows commercial use without asking permission.
Reflection built it mainly for writing code and for agent work, meaning tasks where the AI takes a string of actions by itself rather than just answering a question.
Its own results put it close to last generation's leading Chinese open model while using about a quarter to a third of the computing power to get there and still behind the strongest open models available today.
Those are the company's own numbers, and nobody outside has checked them.
I'd like Tim's view on whether this is something to act on or just a date for the diary.
And for now, it's a promise because the files aren't out and the scores are self-reported.
So I'd treat it as a date for the diary, not a buying decision.
Where it really matters is for organizations whose rules say no Chinese-made models because until now, most of the strongest open options came from China.
If that's you, get a test ready on your own coding jobs so you can run it the week the files land and compare it against whatever you use today.
Isn't the American label a bit of a marketing line, though? A model doesn't care where it was trained.
The model doesn't care, but plenty of procurement and legal teams do.
And for them, the country of origin is a written rule rather than a matter of taste.
That's the real market for Beam.
If your policy has no such rule, judge it purely on how it does with your own work once it's out and ignore the flag on the box.
So a diary note for most people and a proper test plan if origin is on your rule book.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.