Master the captain (@mollicksoif9) on X

What it claims

Reflection AI has released the full weights of Beam on Hugging Face under Apache 2.0. Beam is a 501B-parameter mixture-of-experts agentic coding model with 23B active parameters and a 1M-token context. The post says Beam's selling point is token efficiency, not size. It was RL-trained with a penalty on every wasted token, and the post claims this gives roughly 3-4x less inference compute than GLM-5.2. The post also reports 80.9 on SWE-bench Verified.

How it works

The post gives only two training details:

• Reward: the RL objective penalizes wasted tokens.
• Compute: training ran on 10.5K NVIDIA GB300 GPUs.

It does not describe the reward formulation, data, or MoE routing. It frames cost as price per token × tokens used. On that view, a model that uses fewer tokens can be cheaper overall even at a higher per-token price. That is the comparison to run against your current model, not list price alone.

How to use it

• Weights: full weights on Hugging Face, Apache 2.0. The post gives no repo name.
• Serving: launch support on AWS SageMaker, vLLM and SGLang. The post lists 40+ launch partners but does not name them.
• Hardware: all 501B parameters must sit in GPU memory, about 500GB at 8-bit. The low active-parameter count speeds up compute but does not reduce memory. The post expects most developers to use a hosted API first.
• No commands or model IDs are given.

Limits

• Source: this is a promotional social post from a third-party account, not a model card or paper.
• SWE-bench score: the 80.9 comes with no harness, scaffold, attempt count or baseline models, so it cannot be compared directly with other reported scores.
• Efficiency claim: the 3-4x figure gives no task set, measurement method or setting. The post says independent benchmarks still need to confirm it.
• Quantization: cheaper deployment depends on community quants that the post says do not exist yet.
• Context: the post presents Beam as a US competitor to Qwen, Kimi and DeepSeek, built by two ex-DeepMind founders with NVIDIA backing. This is context for the claims, not evidence for them.
• Before adopting: check the Hugging Face model card for the licence text, the exact checkpoint and the evaluation setup.