Reflection released the full weights for Beam, its 501-billion-parameter mixture-of-experts model, on Hugging Face under the Apache 2.0 license on October 16, 20261,3. The release converts a two-week stretch of benchmark claims and waitlist access into downloadable, forkable artifacts, backed by day-zero hosting from Baseten and a partner ecosystem Reflection pegs at more than 40 integrations.
Architecture and training pipeline
Beam's mixture-of-experts architecture carries 501 billion total parameters with 23 billion active per token and a context window of up to one million tokens. Reflection said it pretrained the model on 23.8 trillion tokens and refined it through a reinforcement learning run that generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks. The final RL run peaked at 883,006 vCPUs, spanning 83% of allocatable capacity across 23 clusters on Google Cloud and Amazon Web Services. Reflection said longer RL rollouts let Beam practice multi-step problems, with a length penalty applied during training to discourage unnecessary tokens.
The company also described custom reliability optimizations that addressed training-inference mismatches and data-quality errors, including a system that reverts a training run to an earlier checkpoint rather than restarting from scratch2. Reflection said these measures enabled it to train Beam with practically zero cluster downtime.
Benchmarks and release artifacts
Reflection reported scores of 80.9% on SWE-bench Verified, 97.8% on AIME 2026 competition mathematics, 90.5% on GPQA Diamond, 80.1% on Terminal-Bench 2.1, and 78.7% on MCP Atlas. Weights shipped in BF16, FP8, and NVFP4 precision formats alongside a speculative-decoding drafter model called DFlash2, a tokenizer, and a reference implementation. The release also includes a model card with final benchmarks and the Beam Safety Technical Report covering safety methodology and evaluations.
Alongside the weights, Reflection open-sourced Mirror CLI, a terminal agent harness trained alongside Beam that pairs a Python frontend with a Rust agent core and supports resumable conversations, parallel agents in separate worktrees, and MCP connections. Reflection said it plans to release self-distillation recipes in the coming days and will adopt OpenEnv for future runs.
Platform and partner breadth
Baseten confirmed day-zero support through Baseten Model API and Dedicated Inference, listing per-million-token pricing of $1.30 for input, $4.30 for cache input, and $0.13 for output. Reflection named Amazon Web Services, Cloudflare, CoreWeave, Databricks, Fireworks AI, Google Cloud, LM Studio, Microsoft Azure, Modal, Nebius Group NV, Nvidia Corp. via NVIDIA NIM, Ollama, OpenRouter, Oracle, Together AI, Vercel, and Weights & Biases as launch platforms. Dell and HPE deliver Beam behind the firewall for on-premises, sovereign, and air-gapped environments. Native day-zero framework support spans Transformers from Hugging Face Inc., vLLM from Inferact, SGLang from RadixArk, Unsloth, Axolotl, TokenSpeed from the LightSeek Foundation, and Dynamo from Nvidia Corp..
Reflection also opened a public beta of its first-party API at platform.reflection.ai, which it said is free to start. CoreWeave confirmed availability spanning training, supervised fine-tuning in partnership with Axolotl, and serverless inference[2].
ANALYSIS The Apache 2.0 license choice removes the restrictive-license friction that has slowed enterprise adoption of other open-weight models, while the breadth of launch partners across cloud, edge, and on-premises channels positions Beam as infrastructure rather than a standalone product. Whether the benchmark numbers hold under independent evaluation will determine how much of that partner momentum converts to production workloads.