Reflection released the weights for Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license on October 16, 2026, and Baseten brought the 501B-parameter mixture-of-experts model to its platform with day-0 support through Model APIs and Dedicated Inference.
Reflection first introduced Beam on October 5, 2026, while the model was undergoing final red-teaming and evaluations, with early access limited to a waitlist and weights promised later in the month. The company said it pretrained Beam on 23.8 trillion tokens and refined it through a reinforcement learning run that generated more than 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks.
Weights, Documentation, and Platform Availability
In its availability announcement, Reflection said the release includes a model card with final benchmarks and a Beam Safety Technical Report covering the model’s complete safety methodology and evaluations. The company described the launch in the availability announcement as an ecosystem of more than 40 partners and integrations forming what it calls an “open operating system for agents,” with the model as the kernel and reference agentic integrations, deployment containers, inference integrations, and customization pipelines open above the infrastructure layer.
Reflection also opened a public beta of its first-party API at platform.reflection.ai, which it said is free to start. Launch platforms named by the company include Amazon Web Services, Baseten, Cloudflare, CoreWeave, Databricks, Fireworks AI, Google Cloud, LM Studio, Microsoft Azure, Modal, Nebius, NVIDIA NIMs, Ollama, OpenRouter, Oracle, Together AI, Vercel, and Weights & Biases, among others. For on-premises, sovereign, and air-gapped environments, Dell and HPE deliver Beam behind the firewall.
Precision Formats and Framework Support
Reflection released the weights in BF16, FP8, and NVFP4 precision formats along with a speculative decoding drafter model. The company said it worked with partners to optimize serving configurations and develop fused kernels for lower latency and higher throughput, steps it said reduce memory requirements and accelerate token generation out of the box.
Beam has native day-zero support in Transformers from Hugging Face, vLLM from Inferact, SGLang from RadixArk, Unsloth, Axolotl, TokenSpeed from the LightSeek Foundation, and Dynamo from NVIDIA. Reflection said developers can download and self-host the model, evaluate it, customize it for their own domains, and connect it to agentic systems through Reflection and its partners.
Day-0 Support and Pricing on Baseten
Baseten announced day-0 support for Beam on October 16, 2026, making it available through Baseten Model APIs and Dedicated Inference. Baseten’s model library describes Beam as a text-only model developed by Reflection under the Apache 2.0 license and lists Model API pricing per one million tokens of $1.30 for input, $4.30 for cache input, and $0.13 for output.
Beam is an autoregressive transformer that accepts text inputs and generates text outputs. Its mixture-of-experts architecture carries 501 billion total parameters with 23 billion active parameters per token and a context window of up to one million tokens.
Benchmark results from Reflection’s evaluations, cited in Baseten’s launch post, report Beam at 90.5% on GPQA Diamond, which measures graduate-level scientific reasoning; 80.9% on SWE-bench Verified, covering resolution of real software issues; 97.8% on AIME 2026 competition mathematics; 80.1% on Terminal-Bench 2.1 for completing tasks in a command-line environment; and 78.7% on MCP Atlas for using real-world tools via MCP servers. Baseten said Reflection trained Beam to solve coding and agentic tasks while using fewer tokens, with longer reinforcement learning rollouts letting the model practice multi-step problems and a length penalty applied during training to discourage unnecessary tokens. Users can adjust Beam’s reasoning effort parameter: lower settings favor shorter responses, while higher settings allow more reasoning on demanding tasks.
Mirror CLI and Customization
Reflection also open-sourced Mirror CLI, a terminal agent harness trained alongside Beam. Mirror pairs a Python frontend covering prompts, skills, and developer tooling with a Rust agent core that drives the model, and it supports resumable conversations, parallel agents in separate worktrees, project instructions, MCP connections, and Python extensions. Reflection said it initially built Mirror for internal engineering use, with isolated workspaces, background sessions, and integrations into Slack, Linear, and GitHub, before evolving it into a tool it could open source.
Beam has been tested with Pi from Earendil, OpenCode from Anomaly, and Hermes from Nous Research across coding and knowledge-work tasks, including reading repositories, editing files, and running commands.
For customization, Reflection partnered with NVIDIA on NeMo RL, AutoModel, and Megatron; RadixArk on Miles; Hugging Face on TRL; Anyscale on SkyRL; and Axolotl to build fine-tuning and reinforcement learning workflows. The company said its largest RL runs for Beam ran roughly 1.3 billion sandboxes across training, evaluation, and ablations, with its sandbox fleet, which ran evaluations and other RL experiments alongside the final run, peaking at 883,006 vCPUs, 83% of allocatable capacity, across 23 clusters on GCP and AWS. Reflection said it plans to release self-distillation recipes in the coming days and will adopt OpenEnv, an open environment standard, for future runs. Baseten describes Beam as the first in a new family of models from Reflection, with multilingual and multimodal capabilities planned for future releases.