Reflection unveils Beam to counter Chinese AI efficiency

Reflection AI officially unveiled Beam on Monday.

The Brooklyn-based startup claims its first frontier, open-weight model matches leading Chinese open models on advanced reasoning benchmarks. It achieves this at dramatically lower costs. This release aims to heat up the race to build a Western answer to DeepSeek, Qwen, and Z.ai.

Beam is a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and features a 1 million token context window. According to the company's blog post, the model was trained using high-compute reinforcement learning to be effective at reasoning, coding, and agentic tasks. Reflection states that Beam uses "3-4x less inference compute" than rivals to achieve similar scores on advanced reasoning benchmarks.

The startup positions Beam against closed labs like Anthropic and OpenAI, as well as popular open models from Chinese developers. Its most direct U.S. rival might be Inkling, the open model from Mira Murati's Thinking Machines Lab released in July. Reflection's own benchmarks show that Beam outscores Inkling on four coding tests where both report results, though Inkling is a multimodal model while Beam is text-only. TechCrunch reported that the announcement confirms prior reporting from Axios that the startup was close to a launch.

Reflection was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. Its last round valued the company at a $25 billion pre-money valuation. This summer, the startup signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia's GB300 chips through 2029. The company is aiming Beam at enterprises and sovereign nations, pitching a product that would let institutions build their own customized, local AI systems by training Reflection's models on their own proprietary data.

The Efficiency Argument

Reflection argues that efficiency is the key differentiator in the current market. On advanced reasoning benchmarks, the company says Beam scores on par with Z.ai's GLM-5.2. However, it claims to achieve these results with significantly lower energy and hardware requirements. This is a direct challenge to the trend of larger, more expensive models dominating the frontier. The company describes Beam as a "workhorse model" for enterprises, the public sector, and developers who need high performance without the associated overhead.

The model's capabilities come from major investments in both pretraining and reinforcement learning. Reflection stated that its high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training. This infrastructure heavy approach is typical of frontier model development, but the focus on inference efficiency is a strategic pivot. By reducing the compute cost per token, Reflection hopes to make advanced AI accessible to a broader range of organizations that cannot afford the massive energy bills of current flagship models.

Competitive Landscape

The open-weight model space is becoming increasingly crowded. Chinese labs have dominated this sector with models that offer high performance at low cost. Reflection's entry adds a significant Western player to this specific niche. The company is not just competing on raw capability but on the economic viability of deploying these models. This is a vital distinction for enterprises that need to manage their IT budgets while adopting AI technologies. The "AI factory" vision championed by Nvidia CEO Jensen Huang, whose company backs Reflection, aligns with this goal. Nvidia's GPUs would power these systems, creating a synergistic ecosystem that benefits both the model provider and the hardware vendor.

According to BleepingComputer, OpenAI is also expanding its monetization strategies, planning to show visual ads in ChatGPT while users generate images. This highlights the broader industry trend of finding new revenue streams beyond subscriptions. While OpenAI focuses on consumer-facing ads, Reflection is targeting enterprise infrastructure. Both approaches reflect the financial pressures on AI companies to sustain their operations. The cost of training and serving large models is immense, and finding efficient ways to generate revenue is a top priority for all major players.

Technical Details and Benchmarks

Beam's performance across a range of coding, agentic, reasoning, and STEM benchmarks is detailed in the company's technical documentation. On the SWE Bench Pro v2-Hard benchmark, Beam scores 77.2, compared to 56.9 for Inkling and 84.3 for GLM 5.3. In the AIME 2026 reasoning benchmark, Beam achieves a score of 97.8. These numbers suggest that Beam is competitive with some of the top models in the field. However, the company acknowledges that frontier open models like Kimi K3 remain ahead on raw capability. Beam's advantage lies in its efficiency at inference time, which can be a decisive factor for large-scale deployments.

The model is undergoing final red-teaming and evaluations before the full release of weights, technical report, and model card. Early access is available for sign-ups. This phased approach allows the company to gather feedback and address any potential issues before the general public release. The focus on coding and agentic performance suggests that Reflection sees these areas as the most valuable applications for its model. By targeting these specific use cases, Reflection can demonstrate the practical benefits of its efficiency claims in real-world scenarios.

Market Implications

The release of Beam has broader implications for the AI market. It signals that Western companies are willing to compete on cost and efficiency, not just raw power. This could lead to a more diverse ecosystem of AI models, with different players catering to different needs and budgets. Enterprises may find it easier to adopt AI technologies if they can choose from a wider range of options that balance performance and cost. The open-weight nature of Beam also means that developers can inspect and modify the model, which can increase trust and facilitate customization. This is a significant advantage over closed models, which are often seen as black boxes.

As the AI industry continues to evolve, the focus on efficiency and cost-effectiveness is likely to become even more important. The energy costs of training and serving large models are a significant concern for both companies and the environment. By developing models that are more efficient, companies like Reflection can contribute to a more sustainable future for AI. The release of Beam is a step in that direction, and it may inspire other companies to follow suit. The coming months will be interesting to watch as the competitive landscape continues to shift.

Reflection's Beam represents a strategic move in the AI race, combining high performance with low cost. By targeting enterprises and sovereign nations, the company is aiming to build a strong foothold in the infrastructure sector. The support from major backers like Nvidia and Sequoia Capital indicates confidence in the model's potential. As the industry matures, the ability to deliver value for money will be a key differentiator. Beam is positioned to be a significant player in this evolving market, offering a viable alternative to both closed and other open-weight models.