Inside this story Context & sources

This article links to 2 sources. Source links are provided below so you can inspect the evidence yourself.

  • Beam’s public weights release is dated 16 October.
  • Modal separately announces its own hosted option.
  • Model size, format and deployment belong in the same evaluation record.

From announcement to downloadable weights

Reflection released Beam’s model weights on 16 October, opening a new route to running its coding and reasoning model. The company says developers can download and host it themselves, use Reflection’s API or access partner services. Beam was introduced earlier in October; the new development is the public weights release and its accompanying deployment options. [1]

That makes the hosting choice a useful place to begin. Our reading is that access to a model and a working system are separate milestones. Before comparing answers, a team needs to know which version it is using, where that version runs and how the result will be checked. A release announcement supplies options; the team still has to choose a concrete setup.

A partner confirms one route

Modal separately announced Beam as a Dedicated Endpoint offering on 16 October. Its notice lists 501 billion total parameters, 23 billion active per token and a context window of one million tokens. Reflection also describes Beam as a sparse mixture-of-experts model, with 501 billion total parameters and 23 billion active. [1] [2]

The total and active counts describe different aspects of the model. We would not use the smaller number alone to decide whether a particular machine can run it. Likewise, a stated context window is a capacity figure; it does not establish how reliably a model will answer questions about everything placed inside it. Our practical question is what the chosen deployment can actually do on a representative task.

Keep the model and the setup together

Reflection says it is releasing the weights in BF16, FP8 and NVFP4 numerical formats, alongside deployment integrations. Those format labels belong in any evaluation record. Modal’s announcement confirms its own hosting route; as a launch partner, it does not provide an independent test of Reflection’s wider performance claims. We have not run Beam or measured either service. [1] [2]

For a South African software team exploring the release, we would begin with a small public demonstration project and one well-defined change. Write down what should happen, keep the initial files, and inspect the proposed change against that expectation. Record the model format or endpoint, the instructions and the result together. This is a suggested evaluation, not a trial we have performed or evidence of local adoption.

The useful result is a repeatable one

A successful first answer would be a reason to inspect more closely, rather than a complete verdict. Repeat the task and note where a person has to correct or explain something. Those observations would make a later comparison more useful than an isolated impressive response.

Beam’s release adds another model that teams can examine through several deployment routes. The immediate value, in our view, is the ability to turn an announcement into a specific, documented experiment. Whether it improves a particular workflow remains a question for that experiment, with the setup and the human review kept in view.

Sources & further reading

1. Reflection’s 16 October public Beam weights release

Dated weights release follows the earlier 5 October introduction. Model capabilities and deployment support are Reflection’s statements, not our measured results.

2. Modal’s 16 October Beam Dedicated Endpoint announcement

Separate company confirms its own Dedicated Endpoint offering and lists model size/context. Launch partner with a commercial interest, not independent performance testing.

AI-assisted research and writing, with a separate source-checking pass by the same Codex assistant. Reflection supplies the model release; Modal independently announces its own hosting offering as a launch partner. Neither source establishes independently tested performance here. Practical evaluation suggestions are our interpretation, not an experiment we ran. No local deployment, measured benefit or hardware fit verified.

THE OPEN TABStay curious. Keep reading.