Reflection AI launches Beam: 501 billion parameters, open weights in October
The US company Reflection AI has launched Beam, the company's first frontier open-weight model. The model's specifications are impressive – 501 billion total parameters, 23 billion active per token, and a one-million-token context window – but the real bet is not raw capacity. Reflection claims that Beam delivers results on par with the best Chinese open models while using "3–4x less inference compute," a figure that for now comes exclusively from the company itself and has not been independently verified. Source: TechCrunch.
What Beam Actually Is
Beam is a text-only mixture-of-experts model (MoE), trained with compute-intensive reinforcement learning for reasoning, coding, and agentic tasks. According to Reflection, as reported by TechCrunch, the model was pretrained on 23.8 trillion tokens and has a one-million-token context window. For comparison, Z.ai's GLM-5.2 has roughly 744 billion total parameters with 40 billion active.
It is the low proportion of active parameters – 23 billion out of 501 billion – that makes the efficiency argument possible: only a fraction of the model runs per generated token. The reinforcement learning phase reportedly ran on 10,500 Nvidia GB300 GPUs for over four weeks, which the company calls one of the largest RL runs any open lab has done, and performance reportedly continued improving until the very end without hitting a ceiling (The Decoder).
The weights are due to be released "later this month" under the Apache 2.0 license, according to The Decoder. That means no third party has been able to evaluate the actual weights as of launch – everything currently known about Beam comes from the company's own figures and benchmark tables.
What Reflection Claims – and What Is Verified
Reflection's core pitch is twofold: Beam is said to score roughly on par with GLM-5.2 on advanced reasoning benchmarks, and to beat today's leading Western open models while using far less inference compute. TechCrunch explicitly stresses that these performance claims are not independently verified.
The compute savings also rest on an estimation method, not measurements. According to how2shout/AInews, which reviewed Reflection's methodology, the company estimates compute cost as roughly twice the number of active parameters multiplied by the mean number of generated tokens – thereby excluding prompt prefill, attention, and serving overhead. Reflection itself describes the comparison as approximate, not as measured inference cost, and uses Artificial Analysis and DataCurve as its source for rivals' figures. The method by construction favors models with fewer active parameters, so the 3–4x figure should be read as a sales pitch with caveats, not a measured result.
Chinese Models Still Lead on Raw Capability
The most revealing part of the launch may be Reflection's own benchmark table. Analysis from how2shout finds that DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 beat Beam on most rows – and that Reflection says so itself: where frontier models like Kimi K3 are ahead on raw capability, Beam's advantage is inference efficiency. According to the same analysis, Beam leads on SWEBench Verified (80.9) and SWEBench Multilingual (78.0).
That means the US open lab is deliberately conceding capability leadership to Chinese competitors and betting everything on the cost advantage. The Decoder headlines Beam as the most capable open-weight model built outside China – a framing that is technically correct but should not be read as Beam leading the open field overall.
Why Efficiency May Be Enough
The efficiency angle is not just marketing. For agentic workloads, where a model may generate tens of thousands of tokens in long chains of tool calls and reasoning steps, inference cost dominates the total economics. Stephanie Zhan, a partner at Sequoia Capital that has invested in Reflection, explained the rationale to New York Times DealBook: what matters most is that the model requires less computing power to complete tasks than Chinese rivals, which is important for compute-intensive applications like agents (NYT).
The New York Times frames the launch more broadly: a successful American open-weight model could threaten the economics of closed-model companies like Anthropic and OpenAI, which finance enormous training costs through API pricing. If open models become "good enough" at a fraction of the cost, the price-to-value ratio of closed APIs comes under pressure. That is the backdrop for the NYT describing Beam as Reflection's first real move in the race.
There is a deeper technical insight behind the bet: Reflection claims that large RL runs on an already capable base model can yield unexpected capabilities. According to how2shout, Beam was trained only on reasoning, software development, and terminal tasks, with no browser tasks in the training mix – but when the model was given web access, it spontaneously began searching, querying other language models, and using OCR APIs to read documents, none of which it was trained to do. The context window was also extended to one million tokens mid-training. Such claims of unplanned capability transfer are interesting, but for now they are only the company's own description and should be verified once the weights are available.
The Company Behind the Model
Reflection AI was founded in 2024 by two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou. The company has, according to TechCrunch citing PitchBook, raised roughly $4.7 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, with its latest round at a $25 billion pre-money valuation. The company has also signed compute deals worth over $7 billion with SpaceX and Nebius, providing access to GB300 hardware through 2029.
In other words: Beam is not a small lab experiment. The capital and compute deals put Reflection in a position to run training runs in the same class as the major laboratories, and the GB300 fleet is the foundation for the large RL phase the company points to as its differentiator.
The Safety Approach
Reflection also describes an unusual safety architecture. According to Benzinga, which reproduces the company's blog post, the company trained a separate safety model with supervised fine-tuning and reinforcement learning, then combined it with the capability-driven RL model through multi-teacher distillation (Benzinga via MSN). The company also says it will publish safety evaluation results and open-source its internal evaluation frameworks. The details of the approach, and how it works in practice, remain to be seen once the documentation and weights become available.
What Is Worth Watching
The decisive question for Beam is simple: do the numbers hold when someone other than Reflection runs the model? Three things will determine that:
- Independent benchmarks once the weights ship under Apache 2.0 later in October – in particular, whether the SWEBench results and the GLM-5.2 parity hold outside the company's own test setup.
- Real inference costs in production, where the prefill, attention, and serving overhead that Reflection's estimate excludes could meaningfully change the actual cost picture.
- Whether efficiency beats capability in the market: Reflection's own table shows Chinese models leading on raw performance. The bet assumes that for agentic workloads, price per completed task matters more than raw benchmark scores.
Until the weights are out, the honest summary is this: Reflection has built a thoughtful, efficiency-oriented model with impressive specifications and serious training infrastructure behind it – but every key figure that distinguishes Beam from its competitors currently comes from the company itself.