Reflection AI released the weights for Beam, a 501-billion-parameter model, under an Apache 2.0 licence on 16 October, according to Amazon Web Services’ deployment guide published that day. The release lets developers run and customise the model while Reflection makes a more particular pitch: it says Beam can complete coding and agent tasks using fewer tokens.
Key points
- Beam has 501 billion parameters in total, but uses 23 billion for each token it processes.
- Reflection trained it with a penalty for unnecessary tokens and reports a score of 80.1% on Terminal-Bench 2.1.
- The weights are available for download, and Beam is also offered through hosted services.
Beam uses 23 billion parameters per token
Beam is a mixture-of-experts model: it holds 501 billion parameters, the adjustable values learned during training, but activates 23 billion for each token, or piece of text, it processes. Think of a workshop stocked for many kinds of jobs, where each job uses selected benches rather than every bench at once. The full set of weights still has to be stored. Amazon Web Services says the model requires multiple GPUs for deployment, while the smaller active portion shapes the computation needed for each token.
Reflection has aimed Beam at coding and agent work, including tasks that involve using tools and acting on their results. The company first previewed the model on 5 October while it was undergoing final red-teaming and evaluations, Unite.AI reported. Nvidia-backed Reflection had been preparing its first open-weight model.
During reinforcement learning, Beam practised tasks and received a penalty when it used unnecessary tokens. Reflection co-founder and chief technology officer Ioannis Antonoglou told VentureBeat that applying such a penalty without enough training can teach a model to avoid solving the task altogether: stopping early is easier than doing the work. Reflection said its four-week run involved more than 100 million rollouts, or attempts at tasks, on 10,500 Nvidia GB300 GPUs.
Editing a file could involve less generated text if Beam completed the edit with fewer tokens, as Reflection says it was trained to do. An unfinished edit would offer no equivalent saving on a completed task. That distinction matters because the charge for generated text depends on both the price of each token and how many tokens the work takes.
Reflection reports 80.1% on Terminal-Bench 2.1
Reflection’s evaluations gave Beam 80.1% on Terminal-Bench 2.1, which tests agent tasks in a command-line environment, according to Unite.AI. The same company evaluations put it at 80.9% on SWE-bench Verified, a test involving software issues, and 78.7% on MCP Atlas, which covers tool use through MCP servers. Those are results for distinct tests: the Terminal-Bench score measures completion of its command-line tasks, not the tokens Beam would use on every coding job.
Reflection is pitching that token use against Chinese open models that US companies have adopted for their lower costs, VentureBeat reported. Antonoglou argued that a buyer ultimately pays for the text a model generates, rather than for a token price in isolation. Reflection’s pitch is that using fewer tokens to complete a task can lower its cost even when each token carries a higher price.
One hosted price is available: Baseten lists Beam output at $0.13 per million tokens, Unite.AI reported. A posted output rate gives a price for generated text; the amount charged for a particular task would also depend on how much text Beam produced. Beam has a reasoning-effort setting that can favour shorter responses or allow more reasoning for demanding tasks.
Beam reaches Hugging Face and SageMaker JumpStart
The weights are available on Hugging Face under Apache 2.0. Reflection released BF16, FP8 and NVFP4 versions, formats with different memory demands and hardware compatibility. For an organisation running Beam itself, selecting a format is only part of the engineering: Amazon Web Services’ deployment guidance calls for capacity to hold the weights and runtime data, followed by a load test to check the estimate.
Reflection also open-sourced Mirror CLI, a terminal agent harness trained alongside Beam. It connects the model to facilities for prompts, developer tooling and running commands. The company said it initially built Mirror for internal engineering use before making it available to others.
Developers who prefer a managed deployment can use an Amazon SageMaker JumpStart real-time endpoint, while Amazon Web Services describes HyperPod Inference routes for those who need control over distributed GPUs and the serving software. At launch, Beam is available through SageMaker JumpStart in US East (N. Virginia) and US East (Ohio).
