News · 2026-10-19

Reflection releases Beam’s weights, but its smallest official package is 294 GB

Reflection released Beam 1.0’s weights under the Apache 2.0 license on October 16, bringing its agent-focused model into public deployment stacks. The model activates 23.4 billion parameters per token but contains 501.3 billion in total. Its smallest official package listed by vLLM is a 294 GB download, with 353 GB of stated minimum GPU memory under that serving recipe.

Key facts

• Beam’s official NVFP4 package is 294 GB on disk; vLLM lists 353 GB of GPU memory for that configuration.
• Reflection published the weights on October 16, 2026, under Apache 2.0.
• The text-only model has 501.3 billion total parameters and a one-million-token context window.
• Primary sources are Reflection’s launch and the vLLM deployment recipe.

The release makes a useful distinction visible: open access and consumer hardware access are separate achievements. A public model can support research, customization and competition even when its normal deployment requires several expensive accelerators. Readers should judge the openness of the release and the affordability of running it as related but different questions.

Beam uses a mixture of experts. Its router selects eight experts from 256 in each expert layer, reducing the amount of calculation used for a token. Imagine a large reference library where a librarian consults only a few shelves for each question. Consulting fewer shelves saves work, but the building still has to store the collection.

That is why the active-parameter headline does not describe the full memory burden. The official model card explains the architecture; the deployment recipe supplies practical capacity figures. vLLM lists about 1,003 GB on disk and 1,204 GB of GPU memory for BF16, versus 503 GB and 604 GB for FP8. These are recipe-specific serving figures, not universal limits for every possible offloading system.

These formats use different numerical precision to store model weights, changing storage needs and serving behavior. The 294 GB package also needs careful naming. Its four-bit format applies to routed experts, while attention, dense components, routing, embeddings and the output layer retain higher precision. The serving recipe includes 20% headroom in its memory estimate for working state. Longer conversations and simultaneous requests add further demands through the key-value cache. Disk capacity alone cannot establish a workable server configuration.

Reflection describes the broader release as an “open operating system for agents.” Its launch connects weights with the Mirror coding harness, containers, serving integrations and customization workflows. The model card describes browsing, tool integration and multi-step software tasks. Those capabilities depend on the surrounding tools and execution loop; a weight file does not itself supply a browser, permissions or a safe place to execute code.

The vLLM engineering post illustrates how deployment work becomes part of model behavior. The serving stack accommodates Beam’s reasoning and tool-call markers and constrains outputs to the expected template. A model can be capable of selecting a tool while an incompatible template prevents the application from using its answer correctly.

A useful outside check comes from Plotly’s October 17 analytics evaluation. Beam answered 36 of 43 questions correctly, tying DeepSeek V4.1 Flash in that run. But the Beam run cost $8.29 compared with $0.46 for DeepSeek. It used about 15 model calls per question and saw very little input caching, while the comparison model benefited from extensive cache reuse.

That result challenges a simple claim that fewer reasoning tokens guarantee a cheaper agent. It does not establish that Beam is inherently more expensive across providers or applications. Plotly tested a narrow, synthetic analytics workload, used a particular hosted pipeline and sells analytics products itself. The comparison is strongest as evidence that calls, caching and orchestration can dominate an end-to-end bill.

Plotly also reports errors involving units, planted patterns and a confounded analysis. Those failures matter because an agent can execute code and still draw the wrong conclusion. Readers can compare the actual package through the Hugging Face release, rather than relying on the active-parameter count or a launch ranking.

The practical significance is a more complete public agent stack with an unusually large deployment footprint. The strongest caveat is that openness does not establish application quality or low operating cost. Beam gives operators another system they can inspect and adapt; whether it is the right system depends on completed-task quality, serving hardware and the full cost of running the work loop.

Primary source, verified: read the paper →

Ground Truth. (2026, October 19). Reflection releases Beam’s weights, but its smallest official package is 294 GB. Ground Truth. https://groundtruth.day/news/reflection-beam-weights-arrive-with-datacenter-memory-needs.html

@misc{groundtruth:reflection-beam-weights-arrive-with-datacenter-memory-needs,
  title  = {Reflection releases Beam’s weights, but its smallest official package is 294 GB},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/reflection-beam-weights-arrive-with-datacenter-memory-needs.html}
}

Topics: open-weights · models · agents · inference · reflection