Reflection Beam is downloadable. What does that put in your hands?

The release changes access, not just the launch date

Reflection introduced Beam on 5 October while final evaluation and red-teaming continued. The availability announcement on 16 October is the concrete change: downloadable BF16, FP8 and NVFP4 versions, a DFlash2 companion model for speculative decoding, and the open-source Mirror coding agent. Its API remains in public beta. Self-distillation recipes, which would help adapt the model using its own outputs, were still promised for the following days.

The published weight repository contains real checkpoint files; the BF16 listing is about 1 TB. Its supplied Apache 2.0 licence permits reuse and modification under its terms. That gives a team a route to choosing its own serving stack rather than relying entirely on a hosted interface. It does not make every partner deployment equivalent. Reflection’s original announcement below establishes what was released.

The independent results favour some tasks much more than others

Artificial Analysis’s 16 October evaluation gives Beam 30 on its Intelligence Index, compared with 39 for Muse Spark Open and 31 for Kimi K2 Horizon. The more revealing split is between individual agent tests: 48.5% on AutomationBench-AA, but 3.0% on Terminal-Bench 4.0. Those are different task sets, not interchangeable measures of one overall success rate.

The same evaluator reports a Coding Agent Index of 28 for Beam in Mirror. Taken together, the results suggest a reason to test specific workflows before choosing an agent. They do not support treating the release’s headline coding claims as a general ranking.

Reflection’s model card reports 80.1 on Terminal-Bench 2.1 for BF16. That figure cannot be set against the independent 4.0 result as though both runs asked the same questions under the same conditions. Benchmark version, precision and the surrounding agent software all matter. A score travels badly when its test setup gets left behind.

Twenty-three billion active parameters still means storing a 501B model

Beam is a mixture-of-experts model: it activates roughly 23.4 billion parameters for each token, from 501.3 billion in total. The smaller number describes the computation selected at each step. The larger one still governs the weights that must be stored and distributed. AWS’s deployment guide explicitly separates those constraints, then adds the memory needed for context and concurrent requests.

A personal hardware test makes the distinction tangible. Stephen J Ronan, whose profile describes a private AI lab and work on FlyCurrent.ai, reports about 238 tokens per second for one stream on two DGX Station GB300 systems, using SGLang, NVFP4 weights and DFlash2 over a 400G link. His single-Station run moved some weights into CPU memory and reached 14 tokens per second. That is a comparison of two serving configurations, not an isolated measurement of the second machine’s contribution.

When Wayne Lowry asked how much speed came from DFlash2 rather than distributing the model, Ronan acknowledged that the test had not separated them. The report is useful evidence of a working setup, with stated hardware and software. It is not an independently replicated speed guarantee or a reason to expect the same result on a laptop.

An open agent gives you choices, including choices you must operate

Mirror is the agent harness: the layer around the model that supplies tools, manages a conversation and decides when an action needs approval. Its overview labels it early beta and describes manual approval as the default. Built-in reading and searching can proceed, while shell commands and file writes pause for permission. Those approvals do not sandbox execution: commands and extensions run with the user’s permissions.

That matters for teams building or maintaining their own sites and internal tools. Inspecting the harness can help explain a result that a model-only comparison misses. It also leaves a team responsible for the environment, permissions and workspaces around it. Open weights and open agent code expand what can be changed; they do not supply an operations team.

There is a concrete regional catch, too. AWS lists JumpStart launch availability in US East, Northern Virginia and Ohio. That announcement does not establish an EU endpoint. A European buyer needs to verify the particular service region and data-handling terms, or assess a self-hosted deployment on infrastructure it controls. The useful question is where this exact configuration runs, with which tools and under whose control.