Reflection Open-Sources 501B Model Beam: Only 23B Parameters Activated per Token, Claims to Match GLM-5.2 with a Quarter of the Compute

On October 16, 2026, Reflection AI released the full weights of Beam, its first open-source model, on Hugging Face under the Apache 2.0 license, along with detailed architecture documentation, a tokenizer, a reference implementation, and a speculative decoding component called DFlash2. The model has 501 billion total parameters (501B), uses a sparse mixture-of-experts (MoE) architecture, activates only 23B parameters per token, has a 1M-token context window, and is positioned for coding and agentic workloads. According to SiliconANGLE, Fortune first reported the release on October 5, and the model only became available for public download on the 16th.

The core numbers in this event are: 501B is the storage scale, and 23B is the inference cost. The design logic of the MoE architecture is to distribute massive parameters across hundreds of "expert" subnetworks, activating only a small subset per token during inference, so actual computation is close to that of a 20-30B dense model. Beam pushes this logic to an extremely low activation rate of 4.6%—of the 501B parameters, only about 23B are used in each forward pass.

Architecture Engineering: How the Memory Wall Was Dismantled

In Beam's architecture design, there is one trade-off with the most direct impact on long-context inference. The model has 52 decoder layers: the first 3 are full-parameter dense layers, and the remaining 49 are MoE layers, each with 256 experts and routing to 8 per token—these 49 MoE layers carry 98.4% of all parameters. The attention mechanism uses a hybrid approach: 43 layers use sliding-window attention (window of 512 tokens, RoPE positional encoding), and 9 layers use global attention (no positional encoding).

This hybrid design brings quantifiable memory benefits. According to the architecture notes published by Hugging Face, Beam's KV cache grows by only 36 KiB per token, whereas a pure global attention design requires 208 KiB/token. At a 1M-token context, BF16 KV cache usage is about 36.08 GiB; with a global design it would be as high as 208 GiB—a difference of about 5.8x.

The model provides checkpoints in three precisions: the BF16 version is about 1TB, while the FP8 and NVFP4 quantized versions are about 280GB, with the latter optimized for NVIDIA Blackwell architecture (GB300 series). Reflection also open-sourced a dedicated Rust-based inference framework, and on the first day of release provided native support for three mainstream inference libraries—transformers, vLLM, and SGLang—as well as supervised fine-tuning (SFT) and reinforcement learning (RL) training recipes.

Training Scale and Benchmark Numbers

According to SiliconANGLE, Beam's pretraining phase used 6,144 NVIDIA GB300 NVL72 GPUs and was completed on a 23.8 trillion-token dataset (including public web data and commercially licensed datasets) in less than four weeks. The reinforcement learning phase was larger: 10,000 GB300 GPUs, with 1.3 billion code generation and tool-calling sandboxes built, also lasting four weeks, with a median failure recovery time of 8 minutes, during which 71 errors requiring manual intervention occurred.

Reflection's self-reported benchmark results are as follows: SWEBench Verified 80.9, Terminal Bench v2.1 80.1, SWE Bench Pro v2-Hard 77.2, AIME 2026 97.8, GPQA Diamond 90.5, and Humanity's Last Exam (without tools) 36.2. According to an evaluation analysis by goldenera.si, some figures listed on Reflection's official blog and Hugging Face page are inconsistent: for the same SWEBench Verified, the blog gives 80.9 while the HF page gives 81.6; for SWE Bench Pro v2-Hard, they are 77.2 and 76.5, respectively. Reflection did not explain the discrepancy.

Reflection says Beam reaches the same level as Z.ai's 744B-parameter model GLM-5.2 on inference benchmarks, but requires only one-quarter to one-third of the latter's inference compute. On Terminal-Bench v2.1, GLM-5.2 reports 81.0 and Beam 80.1, a gap of about 1 percentage point; on LongBench v2, Beam slightly leads GLM-5.2 at 65.5 to 64.0. Reflection's "efficiency advantage" argument holds up in terms of magnitude, but the key methodological detail of how inference compute is measured was not disclosed in the official statement.

What This Means for Different Types of Users

The Apache 2.0 license is the core of this release's commercial impact. Compared with the additional restrictions on commercial use in Meta's LLaMA series, Apache 2.0 allows any organization to fully use, modify, and redistribute the weights without applying for authorization, including for commercial products. Enterprises can deploy on their own hardware or private cloud and keep data within their own infrastructure, which is especially critical for industries with strict data sovereignty constraints, such as finance, healthcare, and government.

For independent developers, Reflection has also set up a Hugging Face ecosystem incentive for Beam: teams that build the best project on the Hub (demo, quantized version, synthetic dataset, RL environment, or benchmark) will receive $10,000 in Hub credits.

For enterprise selection, there are two practical thresholds to assess at this stage. First, the BF16 full-precision version is about 1TB and requires at least eight GPUs with 80GB of memory to load; the FP8/NVFP4 quantized versions drop to about 280GB and are friendlier to Blackwell hardware, but inference quality degradation on other GPUs still needs to be tested in practice. Second, the Rust framework and day-0 library support provided by Reflection lower engineering integration costs, but actual production-grade stability needs to be verified under one's own workloads.

Competitive Coordinates: The Position of US Open-Source Models

SiliconANGLE characterizes Beam as "the first open-source model from a US startup that can match or surpass Chinese open-source models in performance." This characterization has its context: from 2025 to the first half of 2026, the GLM series (Z.ai), Qwen series (Alibaba), and Kimi K3 (Moonshot AI) successively took leading positions in open-source model evaluations, while open weights from the US mainly came from Meta's LLaMA series and Mistral, with relatively limited fully open-source options at the frontier scale.

But comparative data also reveals Beam's limits. On the AA-LCR (Agentic Long-Context Reasoning) benchmark, Beam reports 79.3 while Kimi K3 reports 88.7, a gap of more than 9 percentage points. On the hardest agentic coding tasks, Beam clearly trails Kimi K3, GLM-5.3, and DeepSeek V4.1 Flash. This means Beam is more accurately positioned as matching GLM-5.2 in efficiency-first engineering scenarios, rather than fully catching up with the strongest current open-source models.

Reflection AI was founded in March 2024 by former Google DeepMind researchers Misha Laskin (CEO) and Ioannis Antonoglou (President and CTO). Its latest pre-money valuation was about $25 billion, with cumulative funding of about $4.7 billion, and NVIDIA is a major investor. According to arturmarkus.com, Reflection signed a GPU leasing contract with SpaceX worth about $6.3 billion for training compute. This scale of capital supports the training specification of 10,000 GB300 GPUs and four weeks of reinforcement learning, but it also means Reflection needs to build an open-source community ecosystem and pursue subsequent commercialization to justify its valuation logic.