Reflection Beam: What the 501B Open-Weight Model Actually Is, and What Is Still Missing

Reflection Beam is the first language model from Reflection, the New York lab founded in 2024 by two former DeepMind researchers, Misha Laskin and Ioannis Antonoglou, and one whose stated mission is to "make intelligence open and accessible to all." Beam is a sparse mixture-of-experts model with 501 billion total parameters that wakes 23 billion of them per token, announced on 5 October 2026 and pitched at coding, reasoning and agentic work. It is also, as of 13 October 2026, only half-released: the announcement calls it an open-weight model, the announcement promises Apache 2.0 weights, and the weights are not out.

That gap is the thing to hold on to if you came here from a search for the model's name and want to know whether you can do anything with it. The short answer is: not via weights, not yet, and not on any price list. The long answer is the rest of this page — what Beam is, what Reflection has committed to, what the 501B figure in circulation actually refers to, and what the public record does and does not support. Every figure below is attributed to a source, and none of it has been independently reproduced.

The short version

• What it is — a sparse Mixture-of-Experts text model, 501B total parameters, 23B active per token, built for coding, reasoning and agentic workloads.

• Who publishes it — Reflection, founded in 2024 by Misha Laskin and Ioannis Antonoglou and headquartered in New York. Beam is its first model.

• Announced — 5 October 2026. The benchmark tables in the announcement were revised on 8 October 2026.

• Weights — not released. Reflection says it "will release the weights under an Apache 2.0 license" and that the weights, technical report, model card and developer artefacts are coming "later this month." Neither had shipped as of 13 October 2026.

• Access today — a preview for "a limited set of users," via a waitlist at platform.reflection.ai. No self-serve path, no published price, no download.

• Independent evaluation — none. No third-party harness has scored Beam, and there is no entry for it on the Artificial Analysis leaderboard.

What Reflection says Beam is

Beam's shape is the headline. Reflection describes it as a "sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active" — an activation ratio of roughly 4.6 per cent, which is aggressive even by the standards of the current open-weight field. The architecture notes in the launch post name interleaved local and global attention, fine-grained routed experts, a controlled residual stream, SandwichNorm, elementwise attention gating and FP32 residual accumulation, with auxiliary-loss-free load balancing plus cosine decay of the expert-bias updates.

None of that is exotic on its own. What it adds up to is a serving profile: 23 billion active parameters is a model that can plausibly run at interactive latency on a comparatively small number of GPUs, which is a different product from a 501-billion-parameter dense model even though the number on the card is the same. That is why Reflection can talk about frontier-adjacent reasoning without talking about a data-centre-scale deployment — and why the eventual weight release matters more than the beta key.

Is "501B" the right number? Yes — and it has not moved

This is worth stating plainly because the figure circulates in a few places and readers reasonably wonder whether it is current. The 501B total / 23B active split is what Reflection's own launch post says, it is what the reflection.ai homepage headline says, and it is unchanged as of 13 October 2026. The one update Reflection has published since the launch is a note reading "Update 10/8/26: We've updated Beam's results, and the tables above reflects the latest numbers" — a revision of the benchmark tables, not of the architecture or the parameter count.

What the parameter count does not have is a model card behind it. Reflection has said a model card is coming; it has not said when beyond "later this month," and there is no repository public under Reflection's organisation on Hugging Face as of today ("None public yet" is the literal string on the organisation page). So 501B is a vendor-stated figure, correct as far as anyone can check, and not yet checkable against weights.

Context, modalities and maximum output

Here the vendor material is thinner than the benchmark tables, and it is worth being precise about which number is which, because the two context figures in circulation mean different things.

• Effective context — midtraining "extends Beam's effective context length to 1M tokens." That is a statement about what the model was trained to handle.

• RL rollout context — the reinforcement-learning stage used "a maximum context length of 256K tokens." That is a statement about the training configuration.

• Serving limit — not published. The launch post does not give an API context window, and the endpoint sits behind a waitlist, so there is no second document to check it against. A reader who wants an API context figure today does not have one.

• Maximum output — not published. Reflection does not state one.

• Modalities — text in, text out. The launch post is explicit: "Although Beam is text-only," the model handles other modalities when they are represented as text. There is no image, audio or video input, and the name gives no reason to expect otherwise.

The distinction between the 1M training figure and an unpublished serving limit is the kind of thing that gets flattened in second-hand coverage, and it matters: a training-time capability is not a promise about what an endpoint will accept.

Weights, licence and the "first open model" claim

Reflection's positioning is that Beam is a genuinely open model rather than an API with a marketing adjective attached. The commitment is specific: "we will release the weights under an Apache 2.0 license," along with "documentation and the full stack for running, evaluating, and fine-tuning the model." A separate line in the same post says the weights, technical report, model card and developer artefacts will follow "later this month."

Two things are true at once here. Apache 2.0 is a real, permissive, commercially usable licence — if shipped as described, that is a meaningfully stronger position than a research-only or custom community licence. And a licence promise is not a licence. As of 13 October 2026 the weights have not shipped, so what exists is a stated intention, not a state of the world. The launch post also describes the current version as a preview for a limited set of users, and says the model is in "final red-teaming and evaluations."

The demand signal behind this page is worth one sentence, because it explains why the question keeps getting asked: people are searching the model's name together with the names of other labs' models — Beam plus a Chinese open-weight house, Beam plus a large Chinese open-weight model — which is speculation about what Beam derives from. Reflection documents no basis model. The launch post describes pretraining that ran from scratch: 23.8 trillion tokens across 6,144 GB300 NVL72 GPUs in under four weeks, followed by an RL stage using 10,500 GB300s for four weeks, more than 100 million rollouts and roughly 1.3 billion sandboxes. There is no base checkpoint named anywhere in it. If you have read that Beam is a fine-tune of something else, the vendor's own material does not support it.

What Reflection claims, and what anyone else has measured

Everything in this section is vendor-reported. The launch post's benchmark tables were revised on 8 October 2026 and no third party has reproduced a single figure in them.

• Agentic coding — SWE-Bench Verified 80.9, SWE-Bench Pro v2-Hard 77.2, SWE-Bench Pro v1 65.5, Terminal-Bench v2.1 80.1, DeepSWE v1.1 44.4, SWE-Bench Multilingual 78.0, SWE Atlas Codebase QnA 34.6.

• Reasoning — AIME 2026 97.8, GPQA Diamond 90.5, HLE without tools 36.2, SciCode 49.7, CriPT AA 16.3.

• Tool use and search — MCP Atlas 78.7, BrowseComp 77.4, DeepSearchQA 80.1, AutomationBench 37.0, tau3 banking 38.0.

• General and long context — AA-LCR 79.3, LongBench v2 65.5, IFBench 79.7, AA Omniscience 13.0.

Reflection's own framing of those numbers is careful, and worth repeating rather than upgrading: it positions Beam as "competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max," says Beam matches GLM-5.2 on reasoning while using "3–4× less inference compute," and acknowledges that Kimi K3 still leads on raw capability. That efficiency claim is supported in the post by a chart plotting reasoning score against estimated inference FLOPs, where FLOPs are approximated as roughly twice the active parameter count times generated tokens. The caption says so, and the approximation deliberately excludes prefill, attention and serving overhead. Reflection labels the result "an approximate compute comparison rather than measured inference cost."

Independently: nothing. There is no Artificial Analysis row for Beam, no arena placement, and no third-party harness result we could find. Where Beam and its comparison set have not been benchmarked head to head at matched reasoning effort and matched harness — which is the case for every pair — this page does not imply a ranking. The tables above are one lab's numbers about its own model.

Can you use it, and what would it cost?

Not on published terms. Access is a waitlist: Reflection's post says early access is for a limited set of users, and the platform console sits behind a sign-up wall. There is no published price anywhere on Reflection's own properties — not in the launch post, not on the homepage, not in anything a reader can reach without an account. Any dollar figure you see attached to Reflection Beam right now is somebody's estimate, not a list price.

Reflection does describe the endpoint surface at a high level: an OpenAI-compatible API with a reasoning effort parameter, where "lower settings favor shorter responses, while higher settings allow longer reasoning." The values that parameter accepts are not published in the launch material, and since the model card is still forthcoming there is no second source to check them against.

For readers who are here because they want an efficient reasoning model to call this week, the practical move is the one that does not depend on Beam shipping anything. The models Beam is measured against — GLM 5.2, Kimi K3 and Qwen 3.8-Max — are all live on OrcaRouter today, on one OpenAI-compatible key, with provider list pricing passed through at 0% markup: GLM 5.2 at $1.40 per million input tokens and $4.40 per million output ($1.40/$4.40, read off our own model page on 13 October 2026), Qwen 3.8-Max at $2.00/$6.00, and Kimi K3 at $3.00/$15.00. Because there is no markup, a vendor price cut reaches your bill the same day it is announced rather than at the next contract renewal. If your reason for watching Beam is the efficiency claim rather than the weights, the honest position is that you can measure the alternative today and cannot measure Beam at all.

What to watch, and when the numbers become checkable

Three things would change this page, and all three are datable when they happen. The weights landing on Hugging Face under the stated licence is the first — that is the event that turns a 501B claim into something you can download, quantise, fine-tune and benchmark yourself. The technical report and model card are the second, because they are where a serving context window, a maximum output and the reasoning-effort values would normally be documented; their absence is why three of the spec lines above read "not published" rather than carrying a number. An independent evaluation is the third, and it is the only one that would let anyone outside Reflection say whether 23B active parameters really buys the reasoning scores the launch post claims.

Until then the accurate summary is short. Reflection Beam is a 501B-parameter sparse MoE with 23B active, announced 5 October 2026, text-only, with a waitlisted preview, no published price, no published serving limits, no independent scores, and an Apache 2.0 weight release that has been promised for "later this month" and has not happened. That is not a criticism of the model — it is a description of how far the release has actually got.