Reflection releases Beam open weights under the Apache 2.0 – Unite.AI license

Reflection released weights for Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license on October 16, 2026, and Baseten has ported the 501B parameterized expert-mixing model to its platform with day-0 support via model API and dedicated inference.

Reflection first introduced Beam on October 5, 2026, as the model underwent red-teaming and final evaluations, with early access limited to a waitlist and weights promised later in the month. The company said it pre-trained Beam on 23.8 trillion tokens and fine-tuned it through a reinforcement learning process that generated more than 100 million deployments on 10.5K NVIDIA GB300 GPUs in four weeks.

Weights, documentation and platform availability

In the availability announcement, Reflection said the release includes a model sheet with final benchmarks and a beam safety technical report covering the model’s complete safety assessments and methodology. The company described the launch in the availability announcement as an ecosystem of more than 40 partners and integrations that form what it calls an “open agent operating system,” with the model as reference kernel and agent integrations, deployment containers, inference integrations, and open customization pipelines above the infrastructure layer.

Reflection has also opened a public beta of its proprietary API at platform.reflection.ai, which it says is free to start. The 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-premise, sovereign, and air gap environments, Dell and HPE provide Beam behind the firewall.

Precision formats and framework support

Reflection has released weights in BF16, FP8, and NVFP4 precision formats along with a speculative decoding drafter model. The company said it has worked with partners to optimize service configurations and develop fused kernels for lower latency and higher throughput, steps that reduce memory requirements and speed up token generation out of the box.

Beam has native day-zero support in Hugging Face’s Transformers, Inferact’s vLLM, RadixArk’s SGLang, Unsloth, Axolotl, LightSeek Foundation’s TokenSpeed, and NVIDIA’s Dynamo. Reflection said developers can download and self-host the model, evaluate it, customize it for their own domains, and connect it to agent systems through Reflection and its partners.

Support and pricing from day one on Baseten

Baseten announced day-0 support for Beam on October 16, 2026, making it available via the Baseten Model and Dedicated Inference APIs. Baseten’s template library describes Beam as a text-only template developed by Reflection under the Apache 2.0 license and lists Model API prices for one million tokens as $1.30 per input, $4.30 per cache input, and $0.13 per output.

Beam is an autoregressive transformer that takes text input and generates text output. Its expert mixed 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 assessments, cited in the Baseten launch post, put Beam at 90.5% on GPQA Diamond, which measures graduate-level scientific reasoning; 80.9% on SWE-bench Verified, which covers real-world software problem solving; 97.8% on AIME 2026 mathematics competition; 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 server. Baseten said Reflection trained Beam to solve coding and agent tasks using fewer tokens, with longer reinforcement learning implementations allowing the model to train on 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 answers, while higher settings allow for greater reasoning on challenging tasks.

Mirror CLI and customization

Reflection also open-sourced the Mirror CLI, a terminal agent harness trained alongside Beam. Mirror pairs a Python frontend covering developer prompts, skills, and tools with a Rust agent core that drives the model and supports resumable conversations, parallel agents in separate working trees, project statements, 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 turning it into an open source tool.

Beam was tested with Earendil’s Pi, Anomaly’s OpenCode, and Nous Research’s Hermes for coding tasks and knowledge work, including reading repositories, editing files, and executing commands.

On the customization side, Reflection collaborated with NVIDIA on NeMo RL, AutoModel, and Megatron; RadixArk on Miles; The face hugged on TRL; Anyscale on SkyRL; and Axolotl to create refinement and reinforcement learning workflows. The company said its largest RL runs for Beam ran approximately 1.3 billion sandboxes across training, evaluation, and ablations, with the final run peaking at 883,006 vCPUs, 83% of allocable 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 Reflection templates, with multilingual and multimodal capabilities planned for future releases.