Reflection, a newly established US-based AI laboratory, has officially released its debut model, Beam, under the permissive Apache 2.0 license. Published on Hugging Face on October 16, 2026, Beam is a text-only mixture-of-experts (MoE) architecture designed specifically for coding and agentic tasks. The model features 501 billion total parameters but activates only 23 billion per token, striking a balance between capacity and inference efficiency that is becoming standard for high-performance open weights.

Architecture and Context Efficiency

Beam employs a sparse architecture with 256 experts per MoE layer, routing each token to 8 of them. The model supports a massive 1 million token context window, achieved through a hybrid attention mechanism where most layers utilize a 512-token sliding window, while only 9 of the 52 layers attend to the full context. This design choice significantly reduces memory overhead; at full 1M context, the key-value cache consumes approximately 36 GiB per sequence, compared to the 208 GiB required if every layer were global. The model also includes a dedicated grammar for tool calling, supports five reasoning levels (low, medium, high, xhigh, and max) with 'medium' as the default, and ships with its own open-source harness written in Rust.

Hardware Requirements and Deployment

The release includes three checkpoint formats: BF16 (approx. 1 TB), FP8, and NVFP4. The NVFP4 quantization reduces weights to about 280 GB but requires NVIDIA Blackwell GPUs for native execution, with Reflection suggesting a four-B200 setup for practical long-context usage. Alternative hardware configurations listed include an 8xB200 node, two B300s, or 8xRTX Pro 6000. For those without Blackwell hardware, FP8 is recommended for older GPUs like the H200. Day-one support is available in Transformers, vLLM, and SGLang, while Hugging Face Inference Providers offer access via Baseten, Fireworks, and Together for users lacking local infrastructure.

Benchmarks and Fine-Tuning Potential

Reflection reports strong performance on coding and reasoning benchmarks, citing 81.6% on SWE Bench Verified, 80.1% on Terminal Bench v2.1, 90.6% on GPQA Diamond, and 80.8% on MCP Atlas for the BF16 checkpoint. It is crucial to note these figures are self-reported and not independently verified. Hugging Face also conducted five informal tasks, noting that at high reasoning effort all five passed, while at low effort, one long-context code task and one math task fell short. Fine-tuning is supported via TRL and reinforcement learning with Miles; LoRA supervised fine-tuning fits on a single 8xB300 node, demonstrating improved multilingual reasoning capabilities in held-out tests. Full-parameter GRPO training is more demanding, requiring 96 H200 GPUs in the reference setup, while LoRA GRPO requires 32.

Key Takeaways

  • Beam is a 501B total parameter MoE with 23B active parameters, released under Apache 2.0.
  • It supports a 1M token context window using a hybrid sliding-window/full-attention architecture.
  • NVFP4 quantization requires NVIDIA Blackwell GPUs, reducing memory footprint to ~280 GB.
  • Self-reported benchmarks show 81.6% on SWE Bench Verified and 90.6% on GPQA Diamond.
  • Reflection is offering $10,000 in Hub credits for the best projects built on Beam.

The Bottom Line

Beam is a serious contender for open-weight agentic coding, but its reliance on Blackwell hardware for efficient NVFP4 inference and the lack of independent benchmark verification means early adopters should proceed with caution and verify performance on their own workloads.