Reflection AI put Beam’s weights on Hugging Face on October 16. The 501-billion-parameter model was announced on October 5, with the note that the weights would follow later in the month. They are here now, and they did not arrive alone.
Along with the weights come training recipes: code templates for fine-tuning the model, plus tools that measure how a customization shifts its capabilities and safety guardrails. The weights ship in several versions, including quantized ones that trade output quality for hardware efficiency. There is also a drafter model for speculative decoding, ready-made configurations for several serving engines, and fused kernels that compress multiple inference calculations into a single operation. More recipes are promised, among them one for self-distillation.
Two months of training, essentially without downtime
Beam was trained over two months on several thousand high-end GPUs. At that scale errors are the rule, and they cost model quality. Reflection says its engineers built custom reliability algorithms for this, with particular attention to training-inference mismatches, the bugs that come from small inconsistencies in how a training environment is configured. On top of that sits a system that fixes certain data quality errors by itself: it rolls the run back to an earlier stage and continues from there instead of starting over. The result, the company says, was training with practically zero cluster downtime.
The bet is on the bill per task
Reflection is pitching Beam against the Chinese open models that many US companies already use because they cost less. The argument is not the cheaper token but the token never spent: Beam is meant to finish a task in far fewer tokens than models its size, which lowers the cost of each completed task. In its October 5 post the company said it trained the model with intensive reinforcement learning and penalized it whenever it used unnecessary tokens.
Speaking to VentureBeat, co-founder and CTO Ioannis Antonoglou explained why that penalty is hard to get right. Asked what Beam gave up for its shorter answers, he said: “You don’t give up anything. You just penalize it in terms of how many tokens it can use.” The difficulty lies elsewhere. A team that applies the penalty naively ends up with a model that “will just learn to not solve the task,” because producing fewer tokens is easier than doing the work. The fix, he said, is a very large amount of reinforcement learning, until the model learns to solve the task and stay brief at the same time.
Beam landed a day after Meta’s 816-billion-parameter Muse Spark Open. Meta sells the bigger model on capability, Reflection the smaller one on efficiency. By each company’s own published scores, neither matches the top Chinese models on Terminal-Bench 2.1. That is exactly what Reflection is counting on: as long as performance on agent tasks stays competitive, enterprises should care more about how many tokens a task costs. It is a bet that buyers will start doing their arithmetic in tasks rather than tokens. So far they don’t.