Reflection’s Beam-1.0 Puts a 501B Open-Weight MoE Model Under Apache 2.0

TL;DR

• Reflection released Beam-1.0-501B-A23B, an open-weight mixture-of-experts text model under Apache 2.0, with 501.3B total and 23.4B active parameters.
• Instruction-tuned checkpoints ship in BF16, FP8 and NVFP4, with a 1M-token context window.
• Reflection reports 81.6% on SWEBench Verified and 97.7% on AIME 2026 in BF16, and says Beam matches leading open models using roughly 3x fewer inference-time reasoning tokens.
• Those are the lab’s own figures. The source material includes no independent replication.

Inside Beam-1.0-501B-A23B

Reflection has published Beam-1.0-501B-A23B on Hugging Face. It’s a sparse mixture-of-experts text model with 501.3 billion total parameters and 23.4 billion active ones, spread across 256 experts. The network runs 52 layers: 3 dense and 49 MoE. Vocabulary size is 152k.

The model card puts the pitch plainly: “Beam uses a sparse mixture-of-experts (MoE) architecture with 501B total parameters and 23B active parameters. It supports a context window of 1M tokens and was developed predominantly for English.”

Instruction-tuned checkpoints come in three numeric formats: BF16, FP8 and NVFP4. The license is Apache 2.0. For the BF16 checkpoint, Reflection lists 81.6% on SWEBench Verified and 97.7% on AIME 2026.

Then comes the efficiency claim. Reflection says Beam reaches comparable solution quality to leading open models while using approximately 3x fewer inference-time reasoning tokens. The model is aimed at token-efficient reasoning, software engineering and agentic workflows.

Why 23B Active Parameters Changes the Math

Think of a cathedral pipe organ with 256 sets of pipes. Any single chord sounds only a handful of them, so the music is cheap to play. But the whole instrument still has to fit inside the building. That’s the MoE bargain in one image, and it’s the right way to read Beam’s headline numbers.

Only about 23.4B of the 501.3B parameters do work on a given token. That’s under 5% of the total, which is why a model this large can be pitched as practical to run. Still, the full set of weights has to live somewhere. The three precision formats exist for exactly that reason, since FP8 and NVFP4 checkpoints are built for tighter compute budgets than BF16.

And the 3x claim is the part I’d circle in red. Reasoning models spend tokens thinking before they answer, and for agents that loop through dozens of steps, those tokens pile up fast. If Beam really reaches the same solution quality with a third of the reasoning output, the cost and latency story for coding agents gets interesting. If it doesn’t, it’s a footnote.

So is the claim believable? Maybe. The source gives no named baseline models and no methodology, so “leading open models” is doing a lot of unspecified work. I’d treat the 81.6% SWEBench Verified score the same way: promising, specific, and unverified by anyone outside Reflection so far.

Apache 2.0 deserves its own mention. It’s a permissive license, and that matters more than it sounds when a legal team is deciding whether a 500B-class model can go into a product. Open weights with a clean license remove a whole category of meetings. Worth a lot. Quietly.

Reflection’s First Open-Weight Family

Beam is Reflection’s first open-weight model family, so there’s no earlier release to compare it against. That makes the debut unusually ambitious. Starting at 501B total parameters and a 1M-token context window isn’t a cautious opening move.

Per the model card, the training mix was curated web content, source code and synthetic reinforcement learning environments. That last ingredient fits the agent and software engineering focus. The knowledge cutoff is June 30, 2026, which is useful to know if you plan to ask it about recent libraries or events.

Two limits are worth stating. Beam is a text model, so don’t expect image or audio input. And it was developed predominantly for English, so teams working in other languages should test before assuming parity. The source doesn’t report multilingual results, and I won’t guess at them.

Three Checks Before You Build on Beam

First, watch for independent runs of SWEBench Verified and AIME 2026. A self-reported 81.6% and 97.7% are strong on paper, and outside reruns will show whether they hold up under someone else’s harness.

Second, test the token-efficiency claim on your own tasks. Measure reasoning tokens per solved problem against whatever open model you use today. A 3x gap would show up quickly in your logs, and a smaller one would too.

Third, compare the checkpoints. Reflection’s benchmark figures are listed for BF16, so the FP8 and NVFP4 versions need their own quality checks before you trade accuracy for memory. It’s also worth seeing whether more models join the Beam family, though nothing in the source says they will.