Skip to content
AI.info

The Pulse

Reflection Announces 501-Billion-Parameter Beam and Promises Open Weights Later in October

Reflection announced Beam, a 501-billion-parameter model aimed at coding, reasoning and agentic tasks. The company will offer early access by waitlist before releasing the model weights under an Apache 2.0 license later in October.

Reflection Announces 501-Billion-Parameter Beam and Promises Open Weights Later in October

Reflection announced Beam on October 5, describing it as the company’s first open-weight model and a system built for coding, reasoning and agentic work. The sparse Mixture-of-Experts model has 501 billion total parameters, with 23 billion active per token. But the weights are not available yet: Beam is undergoing final red-teaming and evaluation, and Reflection is inviting users to join an early-access waitlist.

Beam’s weights are promised for later this month

Reflection plans to release the weights under an Apache 2.0 license later in October, alongside a technical report, model card and tools for running, evaluating and fine-tuning the model. The company is also working with distribution partners and open-source libraries. Until those materials arrive, developers can assess the model through the limited early-access program, not download and run its weights.

The distinction matters for a model presented as open-weight: the announcement describes a planned release, not a completed one. Reflection has not yet published the technical report or model card that could give users more detail on training data, evaluation methods and deployment requirements.

Strong coding scores, with gaps on other tests

In its own benchmark table, Reflection reports Beam scoring 80.9 on SWE-bench Verified, 77.2 on SWE-bench Pro v2-Hard and 80.1 on Terminal-Bench 2.1. Those results sit below some listed rivals: DeepSeek V4.1 Flash scores 90.6 on Terminal-Bench 2.1, while GLM 5.3 posts 84.3 on SWE-bench Pro v2-Hard. On the HLE no-tools test, Beam scores 36.2, compared with 40.5 for GLM 5.2.

Beam approaches Qwen 3.8-Max on coding and agentic tasks, Reflection says, while conceding that models such as Kimi K3 remain ahead on raw capability. The company’s table includes results from several other models, but marks many entries as unreported. The figures are Reflection’s published comparisons, rather than independent evaluations in the announcement.

The efficiency comparison is an estimate

Beam reaches reasoning scores comparable to GLM 5.2 while using three to four times less inference compute, according to Reflection. Its comparison estimates generation compute from active parameters and average generated tokens, counting reasoning and final-answer tokens. The estimate excludes prompt prefill, attention operations that vary with context length, and serving overhead, the company says, so it is not a measurement of total deployment cost.

That qualification is central to the launch pitch: Beam’s 501 billion total parameters do not all activate for each token, and Reflection frames lower inference demand as a practical advantage for coding and agent workflows. Whether that advantage holds for real deployments will depend on factors the published estimate leaves out, including serving and prompt-processing costs.

Training combined 23.8 trillion tokens and large-scale reinforcement learning

Reflection pretrained Beam on 23.8 trillion tokens drawn from web material, public sources and proprietary licensed datasets. Separately, its reinforcement-learning run generated more than 100 million rollouts over four weeks on 10,500 NVIDIA GB300 GPUs. The company also used about 1.3 billion sandboxes for training and grading.

Beam can adjust reasoning effort: lower settings favor shorter outputs, while higher settings allow more reasoning on demanding tasks. The model is text-only, though its demonstrations include tasks that use text-represented information from other modalities. For now, the announced release schedule puts public access to the weights after the current waitlist phase, with Apache 2.0 licensing planned for later in October.

Sources

  1. Original source · Blog post Reflection reflection.ai

Related stories

Explore

More articles