Beam
A 501B sparse mixture-of-experts model with 23B active — Apache 2.0, aimed squarely at the Chinese open-weight labs.
An open-weight model built for coding and agent work. It uses a mixture-of-experts design, which means that although it holds 501 billion parameters, only about 23 billion of them fire for any given token — so it costs far less to run than the headline number suggests.
What you build with it
- Self-hosted coding assistants where source code cannot leave the network
- Agent workloads at a volume where per-token API bills have stopped making sense
- Fine-tuning on your own codebase
- A Western-licensed alternative to the leading Chinese open models
What it gets you
- Apache 2.0 — genuinely permissive, no revenue thresholds or use limits
- Compute cost closer to a 23B model than a 501B one
- Context window up to a million tokens
- Parameters
- 501B total · 23B active
- Architecture
- Sparse MoE, text-only
- Context
- Up to 1M tokens
- Licence
- Apache 2.0 (pending release)
- SWE-Bench Verified
- 80.9 (company’s own)
Straight from the source
Why it matters
Apache 2.0 at this scale is rare, and the architecture is the reason it is interesting rather than just large. 501B total parameters with 23B active means you pay memory for the big number and compute for the small one — which is what makes a half-trillion-parameter model something a team can realistically serve.
23B active is the number that matters
In a sparse mixture-of-experts model only a fraction of the network fires for any given token. Beam holds 501B parameters in total but activates about 23B per token.
The practical consequence: you need memory to hold all the weights, but your compute cost per token is closer to a 23B dense model than a 501B one. Memory is the cheaper constraint to buy your way out of.
That is the whole reason MoE has taken over at this scale, and why a number like "501B" tells you much less about serving cost than it used to.
Apache 2.0 is the other headline
Reflection says it will publish the weights under Apache 2.0, together with the technical report, model card, and the full stack for running, evaluating and fine-tuning the model.
Apache 2.0 is genuinely permissive — no revenue thresholds, no field-of-use restrictions, no acceptable-use addendum that can be revised later. That is a materially different proposition from the "open" community licences several large models ship under, and it is the difference between a model you can build a company on and one you can experiment with.
Reflection is positioning Beam explicitly as a Western alternative to leading Chinese open models, and claims parity with Z.ai's GLM-5.2 on advanced reasoning at 3–4× less inference compute.
It is not actually out yet
As this entry was written, Beam had been announced but the weights had not been published — Reflection said they would follow later in October, after final red-teaming.
Access so far is a waitlist for a select group plus a beta API. The benchmark figures (80.9 on SWE-Bench Verified, 80.1 on Terminal Bench v2.1) are the company's own and have not been independently verified.
Worth watching, not yet worth planning around. Check Reflection's blog or Hugging Face for the actual drop before anything depends on it.
Reach for it when
- You need genuinely permissive weights for coding or agent workloads
- MoE serving economics fit your hardware — memory-heavy, compute-light
- You want a Western-licensed alternative to the Chinese open-weight models
Watch out for
- The weights were not published yet — confirm before planning around it
- All benchmark numbers are the company's own and unverified
- Text-only, so no multimodal workloads