Beam: Reflection's 501B open-weight model
Reflection.ai introduced Beam, a sparse 501-billion parameter open-weight Mixture-of-Experts model trained on 23.8 trillion tokens with massive reinforcement learning.
summary is enough The announcement details a massive training scale and architecture, but weights are delayed and benchmark performance trails existing open alternatives.
summary
Reflection.ai introduced Beam, a 501-billion parameter sparse Mixture-of-Experts model with 23 billion active parameters, built for coding, reasoning, and agentic workflows. The model underwent extensive pretraining on 23.8 trillion tokens followed by a reinforcement learning run utilizing 10.5K NVIDIA GB300 GPUs. Although the company plans to release model weights under an Apache 2.0 license later this month, hacker news commenters quickly noted that Beam trails smaller, more efficient open models like DeepSeek v4.1 Flash on benchmarks.
key points
- Beam is a sparse Mixture-of-Experts model featuring 501 billion total parameters and 23 billion active parameters.
- The training pipeline involved 23.8 trillion pretraining tokens and over 100 million reinforcement learning rollouts across 10.5K NVIDIA GB300 GPUs.
- Model weights, technical reports, and developer artifacts are scheduled for release later this month under an Apache 2.0 license.
who should read it
Machine learning engineers and AI researchers tracking open-weight frontier models.
discussion skeptical
consensus
Commenters are heavily skeptical, highlighting that Beam is larger and underperforms compared to existing open-weight models like DeepSeek v4.1 Flash while withholding its weights at launch.
pushback
A few commenters welcomed Beam as a much-needed Western entrant to the open-weight ecosystem, arguing that new competitors drive healthy industry iteration.
I thought it would be interesting to look at some key figures vs another contemporary model in the same weight class (DeepSeek V4.1 Flash)
DS V4.1F Beam LM total params 552B 501B LM active params (prefill) 8B 23B LM active params (decode) 16B 23B N-gram/PLE params 196B 0 Pretrain tokens 45T 28T Disk KV bytes/token (FP4) 890 No information Vision Yes (pretrain) No Weights available Yes (launch day) "This month" Weights licence MIT Apache 2.0
At first blush the benchmarks are impressive, but to paraphrase Linus: "Talk is cheap, show me the weights." :-)
Provides a clear tabular side-by-side comparison of Beam and DeepSeek V4.1 Flash, exposing the parameter, token, and license tradeoffs.
Generated by an LLM from the article and thread. Check the sources before relying on it.