This Monday (October 5), a Brooklyn, New York–based startup Reflection AI Officially Unveils Its First Flagship Open-Source Weight Model BeamDirectly Targeting China’s Open-Source Camp, Represented by GLM-5.2,Qwen3.8,DeepSeek V4 This Two-Year-Old Team—Valued at $25 Billion—Claims Beam Matches GLM-5.2 on Advanced Reasoning Benchmarks While Requiring Only About One-Quarter to One-Third the Inference Compute. The Long-Awaited “Open-Source Answer” for Western Developers Finally Has a Credible Contender.
Beam specifications: A 501-billion-parameter Mixture of Experts (MoE) model
Beam is a bona fide “flagship-grade” open-source model, with the following core specifications:
- Total parameters: 501 billion; active parameters: 23 billionUses a Mixture of Experts (MoE) architecture—only 23 billion parameters are activated per inference, balancing performance and speed
- Pretrained on 23.8 trillion tokensTraining data scale places it among the top tier
- 1-million-token context windowMatches current flagship open-source models—capable of processing an entire novel or large codebase in one go
- Text-only modelDoes not yet support multimodal inputs such as images; positioned as a “workhorse” for reasoning, programming, and agent tasks
For comparison, Zhipu’s GLM-5.2 has approximately 744 billion total parameters and 40 billion active parameters. Beam matches competitors using fewer active parameters, enabled by high-compute reinforcement learning (RL) training—the key technical differentiator among cutting-edge models today
Performance positioning: Benchmarked against GLM-5.2 and directly challenging Inkling
Reflection’s performance claims have not yet undergone independent third-party verification, but the company’s stated targets are highly ambitious: on advanced reasoning benchmarks, Beam’s scoresMatch Z.ai’s GLM-5.2whileSurpass existing Western open-source modelsWhile requiring only “one-third to one-quarter the compute” of its rivals during inference. Reflection calls it a “workhorse model” for enterprises, public-sector organizations, and developers
Within the Western camp, Beam’s most direct competitor is InklingInkling—a multimodal open-source model released in July this year by Thinking Machines Lab, founded by Mira Murati. Reflection’s own benchmarks show Beam outperforming Inkling across four programming tests; however, Inkling is multimodal while Beam is text-only, so their use cases do not fully overlap
| Model | Parameter count | Core Positioning | context |
|---|---|---|---|
| Reflection Beam | 5010B / 230B active | Reasoning · Programming · Agents | 1 million |
| GLM-5.2 | ~7440B / 40B active | Open-source programming flagship | 1 million |
| Qwen3.8 2.4T | 2.4T total parameters | Open-source general-purpose flagship | Ultra-long |
| MiniMax M3 | Million-token open weights | Programming Agent | 1 million |
| Nemotron 3 Ultra | 550B | Strongest open-source model in the U.S. | — |
Who is Reflection AI: A $4.7 billion bet by two DeepMind veterans
Reflection AI was founded in 2024 by two former Google DeepMind researchers. According to PitchBook data, the company has raised approximately $4.7 billionfrom investors including Nvidia, Sequoia Capital, and Lightspeed, with its latest round’s pre-money valuation reaching as high as $25 billion.
The money is going toward compute. This summer, Reflection signed compute contracts totaling over $7 billionwith SpaceX and Nebius, securing supply of Nvidia’s GB300 chips through 2029—a hard requirement for training frontier models and competing for customers against Anthropic and OpenAI.
Reflection’s true ambition extends beyond selling models—to selling an “AI factory”: enabling enterprises and sovereign nations to train customized, localized AI systems using their own private data. This vision aligns closely with Nvidia CEO Jensen Huang’s long-standing “AI factory” concept—and explains why Nvidia backed the company: the more closed models OpenAI sells, the larger Nvidia’s GPU business grows.
What this means for users
Let’s state this clearly first:Beam’s weights are not yet available for downloadReflection states that the weights and full technical details will be released “within this month,” distributed via major cloud providers, neocloud, and open-source libraries. So you cannot run it locally yet—this is primarily an official announcement and preview.
But the signal it sends is clear: the Western open-source camp, having been outperformed by Qwen and GLM for over a year, is finally fielding a heavyweight contender capable of head-to-head competition. For domestic users and developers, this means three things: DeepSeekMore choices
- More optionsOnce the weights are released, enterprises and developers gain another Western open-source base model option, reducing reliance on a single vendor;
- Pricing pressureBeam emphasizes “achieving equivalent performance with less compute,” potentially further lowering inference costs and pressuring domestic models to reduce prices;
- Open-source ecosystem vitalityWith more cutting-edge weights going open source, all users capable of self-hosting ultimately benefit.
Frequently Asked Questions (FAQ)
Is Reflection Beam free? Can I download its weights?
Beam is positioned as an open-weight model; Reflection has announced that the weights and full technical details will be released within this month, accessible via cloud providers, neocloud, and open-source libraries. Downloads are not yet available; the specific license terms will be announced officially.
What is Reflection AI?
Reflection AI is an AI startup founded in 2024 in Brooklyn, New York, co-founded by two former Google DeepMind researchers, with approximately $4.7 billion in cumulative funding from investors including Nvidia and Sequoia Capital, and a latest valuation of around $25 billion.
How does Beam compare with GLM andDeepSeek ?
According to official claims, Beam matches GLM-5.2 on advanced reasoning benchmarks, though this has not yet been independently verified by third parties. Beam is a pure text model focused on reasoning, programming, and agent tasks; many GLM andDeepSeek models already support multimodality. Which performs better requires real-world benchmarking after the weights are publicly released.
What is Beam’s parameter count? What hardware is required to run it?
Beam has 501 billion total parameters and 23 billion active parameters, making it a flagship MoE model; running the full weights requires a multi-GPU high-performance server. Individual developers will mostly rely on quantized versions or API interfaces provided by cloud vendors.
Want to stay updated on open-source model developments? Check out our AI Model Library and Tool Comparison Engineor continue reading:DeepSeek V4 Pro benchmark · GLM 5.3 Evaluation · Atria Dawn 744B open-source model evaluation.