501.3B total / 23.4B active (sparse MoE)
1,048,576 tokens
Apache 2.0
Reflection's first open-weight model, released as an English-predominant, text-only instruction model for software engineering, reasoning, tool use, browsing, and long-horizon agentic workloads. Beam 1.0 is a sparse mixture-of-experts Transformer with 501.3B total parameters and 23.4B activated per token. Its 52-layer language backbone begins with three dense layers followed by 49 MoE layers, each containing 256 routed experts and selecting eight experts per token. The released configuration uses a 5,120-dimensional hidden state, 20,480-dimensional dense feed-forward layers, 2,560-dimensional expert feed-forward layers, 48 query heads, eight key/value heads, 128-dimensional attention heads, an untied 151,936-token vocabulary, and a modified Harmony prompt format. Gated grouped-query attention combines mostly 512-token sliding-window layers with periodic full-attention layers. The native configuration and tokenizer specify 1,048,576 positions, and Reflection lists a June 30, 2026 knowledge cutoff. Beam was pretrained on 23.8 trillion tokens from curated public web data, source code, technical documentation, educational and scientific material, and commercially licensed datasets. Mid-training adds curated code repositories, long-form documents, tool-use trajectories, and long-horizon tasks. Reinforcement learning spans coding, agentic, search, tool-use, and STEM environments generated primarily through synthetic pipelines and supplemented by proprietary vendor and open-source data. Reflection reports more than 100 million high-compute RL rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. Across training, evaluation, and ablations, the company reports approximately 1.3 billion sandbox instances; the peak hour of the final RL run reached 883,006 allocatable vCPUs across 23 GCP and AWS clusters. Reflection positions Beam as reaching comparable solution quality to leading open models while consuming approximately three times fewer inference-time reasoning tokens. Final developer-reported BF16 scores include 43.3 on DeepSWE v1.1, 76.5 on SWE-bench Pro v2-Hard, 66.1 on SWE-bench Pro v1, 80.1 on Terminal-Bench 2.1, 78.3 on SWE-bench Multilingual, 81.6 on SWE-bench Verified, 97.7 on AIME 2026, 36.0 on Humanity's Last Exam without tools, 51.1 with tools, 90.6 on GPQA Diamond, 88.3 on LiveCodeBench, 80.8 on MCP Atlas, 36.1 on tau3-bench Banking, 76.8 on BrowseComp with context management, 80.5 on DeepSearchQA with context management, 67.5 on LongBench v2, and 79.6 on IFBench. Reflection reports broadly similar results from the compressed checkpoints, including several cases where FP8 or NVFP4 slightly exceeds BF16. The release includes 188 BF16 safetensors shards totaling approximately 1.003 TB, 188 FP8 shards totaling approximately 503.3 GB, and 53 NVFP4 shards totaling approximately 293.6 GB. A separate 1.93B-parameter DFlash2 speculative-decoding model uses five layers, a 4,096-token sliding window, and hidden states drawn from five Beam layers to draft blocks of tokens which Beam then verifies. On four NVIDIA B300 GPUs with Beam NVFP4, Reflection reports mean speculative acceptance lengths of 3.93 to 5.60 tokens and greedy-decoding throughput gains ranging from 1.72 to 3.03 times across GSM8K, MATH-500, HumanEval, MBPP, and MT-Bench. Day-zero deployment support includes Transformers, vLLM, SGLang, NVIDIA Dynamo, TokenSpeed, Unsloth, and Axolotl. Reflection additionally released the Mirror CLI coding and agent harness, and reports testing Beam with Pi, OpenCode, and Hermes. The launch includes customization integrations with NeMo RL, AutoModel, Megatron, Miles, TRL, SkyRL, Axolotl, and OpenEnv, with self-distillation recipes announced for a subsequent release. Reflection's Iris Safety Evaluations report an 88.9% harmful-request refusal-or-pivot rate and an 81.5% harmless-request compliance rate for BF16. The Apache-2.0 release explicitly permits commercial and research use, fine-tuning, parameter-efficient adaptation, quantization, merging, distillation, and using model outputs to train other models. Default evaluation sampling uses temperature 0.7 and top-p 0.99.