Beam 501B: 23B active parameters, Apache 2.0 licensed

Reflection, a US-based AI lab, has released Beam, a 501-billion-parameter open-weight model available on Hugging Face. The model represents a shift toward providing organizations with frontier-scale capabilities they can control and customize without relying on proprietary systems.

What changed

Beam is Reflection’s first model release, arriving as a sparse mixture-of-experts architecture with 501 billion total parameters but only 23 billion active per token. The model processes text input and output with a 1-million-token context window under an Apache 2.0 license. Multiple checkpoint formats are available: BF16, FP8, and NVFP4, each designed for different hardware generations.

The model targets coding and agentic workflows specifically. Beam includes a Rust-based inference harness for performance optimization and integrates immediately with transformers, SGLang, and vLLM. Hugging Face offers inference, supervised fine-tuning, and reinforcement learning recipes to help teams deploy and adapt the model.

Beam’s residual stream design differs from standard transformers. Rather than adding sublayer outputs with a fixed weight of 1, Beam scales additions by 1/sqrt(k) at layer k while scaling the existing stream by sqrt((k-1)/k). This keeps the residual stream at constant magnitude across all 52 layers, allowing each sublayer to contribute equally to the final output.

For long-context handling, Beam employs hybrid attention. Most of its 52 layers use sliding-window attention covering only 512 previous tokens; every sixth layer plus the final layer process the full context. This architecture reduces per-token cache size from 208 KiB (if all layers were global) to 36 KiB, allowing a 1-million-token sequence to fit in approximately 36 GB of memory instead of 208 GB.

Why it matters

Open-weight models like Beam provide alternatives that proprietary systems cannot match. Organizations can deploy Beam on their own infrastructure with complete control over customization and data handling. This addresses a critical need for enterprises requiring frontier-scale reasoning for code generation or complex agent tasks while avoiding vendor lock-in or third-party API dependencies.

The sparse mixture-of-experts approach offers practical benefits. Though all 501 billion parameters remain in memory, only 23 billion activate per token. Beam therefore delivers throughput comparable to a 23-billion-parameter model while retaining the reasoning capacity of a much larger system. This efficiency matters for organizations operating at scale or working within cost constraints.

Beam’s emphasis on coding and agentic tasks reflects concrete market demand. Coding requires precise reasoning and tool interaction; agentic systems need sustained reasoning across multi-turn conversations. The model supports five reasoning levels (“low” through “max”), tool calling with dedicated grammar, and reasoning states that persist across turns. These features address specific production workflows.

The 1-million-token context window removes bottlenecks for teams processing large codebases, extended conversations, or complex reasoning chains. Like other models with longer contexts, Beam solves the memory problem through hybrid attention rather than uniform cache scaling.

What to test

Before deploying Beam, validate these aspects:

• Run inference benchmarks on your own hardware and compare actual throughput and latency against reported speeds. Test across BF16, FP8, and NVFP4 checkpoints to determine which precision balances accuracy and performance for your use case.

• Deploy Beam on your target infrastructure and measure memory usage and inference cost. Reflection recommends four B200 GPUs for production workloads with long contexts; verify this works at your expected request volume and sequence lengths.

• Fine-tune or customize Beam on your own data using the provided supervised fine-tuning and reinforcement learning recipes and measure whether training costs justify quality improvements over the base model.

• Compare the Rust-based inference harness against vLLM and SGLang in your environment. Performance characteristics vary by hardware and configuration.

• Evaluate coding and reasoning performance on your specific test cases. Beam achieves 81.6% on SWEBench Verified and 90.6% on GPQA Diamond, but real-world accuracy depends on task characteristics and customization.

• Confirm that Apache 2.0 licensing meets your compliance and commercial requirements before production deployment.

The conclusion

Beam addresses a genuine need for teams requiring frontier-scale coding and reasoning capabilities while maintaining full ownership and control. Its sparse architecture provides practical efficiency, and its open-weight status eliminates licensing complexity. The 1-million-token context and tool-calling support serve real agentic use cases.

However, Beam is not a replacement for proprietary frontier systems in all scenarios. Teams should test performance on their specific tasks before committing to production use. The underlying engineering appears sound, but benchmark results do not always translate directly to production outcomes.

Reflection is offering $10,000 Hub credits to the best projects built with Beam, providing incentives for early adoption and specialized applications. The community’s extensions through quantization, reinforcement learning training, and domain-specific fine-tuning will likely determine Beam’s long-term value as much as the base model itself.

Related reading

• Apertus 1.5: 262K context, image and audio support
• QwenGyre elastic RL framework trains long-horizon agents
• OrcaCyber Zero 1.5 Cybersecurity Model Launches With 1M Context