Reflection releases Beam, a 501B open-weight model under Apache 2.0
Article timeline
The articles in this story. Times are ET.
Sophia Yang@sophiamyangXPickReflection AI releases Beam, an open-weight coding model under Apache 2.0AIReflection AI has released Beam with open weights under the Apache 2.0 license, along with weights, a tokenizer, a reference implementation, and the open-source Mirror CLI coding tool. The model has 501B total parameters with 23B active, a 1M context window, and scores 81.6 on SWE-bench Verified and 80.1 on Terminal-Bench 2.1. The author reports a 1.9B drafter model that speeds up generation 2–3x.
NVIDIA AI@NVIDIAAIOfficialPickReflection AI releases Beam, a 501B MoE open-weight modelAINVIDIA AI welcomes Beam, the first open-weight model from Reflection AI, released under the Apache 2.0 license with a 40+ partner ecosystem. The 501B MoE model was trained on NVIDIA GB300 NVL72 and delivers over 70K tokens/sec/GPU on Blackwell Ultra on day 0. Weights, tokenizer, a reference implementation, BF16, FP8 and NVFP4 weights, a DFlash2 drafter, and the Mirror CLI harness are released.
SGLang@sgl_projectOfficialPickSGLang adds day-0 support for Reflection's open-weight Beam modelAISGLang says its production stack now serves Reflection's Beam, a model built for software engineering and long-horizon agents, with day-0 support. The post describes Beam as delivering frontier-level inference efficiency with more intelligence per token and links a serving cookbook and a blog post for details.
ollama@ollamaOfficialOllama adds Reflection's Beam model to its libraryAIOllama announces a model page for Beam and a blog post about it, both linked from the post. The post says little more, so the summary stays brief.
- ollama@ollamaOfficialPickOllama offers free cloud access to Reflection's open-weight Beam model
AIOllama says it is a day 0 partner for Reflection's first open-weight model, Beam, now free for a limited time on Ollama's cloud. Ollama says the model is hosted in the US with full ZDR, has a 1M token context window, and is designed for software engineering, reasoning, and long-running agentic tasks. Reflection's quoted post says Beam's open weights use the Apache 2.0 license and come in BF16, FP8 and NVFP4 formats, with its Mirror CLI released as open source.
Artificial Analysis@ArtificialAnlysOfficialBeam leads AutomationBench-AA and CritPt in Intelligence Index testsAIArtificial Analysis says Beam scores 48.5% on AutomationBench-AA and 13.1% on CritPt, more than double K2 Horizon's 4.6% and Qwen3.8 27B's 5.4% on those evaluations. Its weakest results are 3.0% on Terminal-Bench 4.0 and 976 Elo on AA-Briefcase.
- Artificial Analysis@ArtificialAnlysOfficialBeam scores 34% on Artificial Analysis Cyber Index, leading US open-weights models
AIArtificial Analysis says Beam scores 34% on its Cyber Index, the highest result among US open-weights models it has tested. The score is driven by token efficiency, placing Beam on the Cyber Index versus output tokens per task frontier alongside GPT-6 Sol (Daybreak Blue).
- Artificial Analysis@ArtificialAnlysOfficialReflection's Beam scores 30 on Artificial Analysis Intelligence Index, second among US open weights models
AIArtificial Analysis says Reflection's first model, Beam, scores 30 on its Intelligence Index, behind only Meta's Muse Spark Open (39) among US open weights models. Beam uses about 26k output tokens per index task and scores 48.5% on AutomationBench-AA, but only 3.0% on Terminal-Bench 4.0. Its hallucination rate is 47%, though its net AA-Omniscience index score of 1.5 is higher than its peers.
vLLM@vllm_projectOfficialvLLM releases a Docker image and serving recipe for BeamAIvLLM announces a Docker image, a serving recipe and a blog post for Beam, a model it says is ready to serve. In a chat test at concurrency 16, 4× B300 NVFP4 TP4 reached 1,769 output tok/s and 8× B300 FP8 TP8 reached 1,300 tok/s. The post notes the two setups use different GPU counts and precisions, so the figures are not an equal-resource comparison.
- vLLM@vllm_projectOfficialvLLM and SkyRL add day-0 RL support for Beam
AIvLLM says SkyRL now supports RL training for Beam with LoRA, BF16 full fine-tuning, and MXFP8 RL, using vLLM rollouts and Reflection's Megatron-Bridge implementation. SkyRL reports a 62% LoRA decode throughput gain from right-sizing expert adapters in vLLM, rising from 276 to 447 tok/s/GPU at 32 concurrent requests on PP2×TP8, though this is a decode result rather than an end-to-end training speedup.
- vLLM@vllm_projectOfficialBeam model activates 8 of 256 experts per token across 52 layers
AIvLLM's post describes Beam as a model with 8 of 256 experts active per token across 52 layers. Its attention alternates a 512-token sliding window with global NoPE attention, and a separate 5-layer drafter supplies speculative tokens.
- vLLM@vllm_projectOfficialvLLM adds Day-0 support for Reflection-Beam-1.0, a 501B-parameter model
AIvLLM says Reflection-Beam-1.0 has Day-0 serving support, with 501B total parameters, 23B active per token, and a 1M-token context window. The post says Beam combines hybrid attention with Harmony-format reasoning and tool use. SkyRL also adds Day-0 RL support for Beam, including LoRA, BF16 full fine-tuning and MXFP8 RL with vLLM rollouts.
Arena.ai@arenaOfficialPickReflection AI's Beam 1.0 ranks #12 among open models on Agent ArenaAIArena says Reflection AI's Beam-1.0-501B-A23B ranks #12 among open-weight models on Agent Arena, with a -3.67% Net Improvement. It ranks #22 among open models on Code Arena: WebDev with 1438 points and is released under Apache 2.0. Arena notes a ±3.84% margin of error on Net Improvement, placing Beam closer to the two Qwen3.8 models above it than its rank suggests.
Hugging Face@huggingfaceOfficialPickReflection releases Beam open weights under Apache 2.0 with 40+ partnersAIHugging Face relays Reflection's announcement that Beam is now available across platforms with open weights under the Apache 2.0 license, in collaboration with more than 40 partners. Reflection says it is releasing weights, tokenizer, and a reference implementation in BF16, FP8, and NVFP4 formats, plus a DFlash2 drafter and the open-source Mirror CLI harness used to build and train Beam. A live stream with the Reflection team on how Beam was built is scheduled to start shortly.
merve@mervenoyannXHugging Face blog explains how to use reflection-beam with inference enginesAIMerve Noyan of Hugging Face shares a blog post on using reflection-beam with different inference engines. The post includes vibe tests and more, according to the announcement.
- merve@mervenoyannXPickBeam, a token-efficient coding model from Reflection, is available with open weights
AIMerve Noyan says Beam is a token-efficient, frontier-scale model for coding agents, released with its own harness. The quoted Reflection post says Beam is available under the Apache 2.0 license with weights, a tokenizer, and a reference implementation. It also lists BF16, FP8, and NVFP4 weights, a DFlash2 drafter, and the open-source Mirror CLI, with support from more than 40 partners.
Julien Chaumond@julien_cXHugging Face releases Beam, linked to the Reflection repositoryAIJulien Chaumond, the Hugging Face co-founder, announces that Beam is out and links to the Reflection page on huggingface.co. The post gives no details on what Beam does, its size, or its benchmark results.
Reflection@reflection_aiOfficialReflection launches Beam, an open operating system built around its modelAIReflection AI launches Beam, which it calls an open operating system, with its own open-source tools and integrations with developer projects. The company says its model is the kernel and everything above the infrastructure layer is open.
- Reflection@reflection_aiOfficialPickReflection releases open-weight Beam model under Apache 2.0 license
AIReflection says Beam is now available across platforms, with open weights under the Apache 2.0 license and support from more than 40 partners. The release includes weights, a tokenizer, and a reference implementation in BF16, FP8, and NVFP4 formats, plus a DFlash2 drafter and the open-source Mirror CLI harness used to build and train Beam. Weights and the model card are on Hugging Face.
- Reflection@reflection_aiOfficialReflection AI says 40+ partners will extend its Beam stack
AIReflection AI says it will show in the coming days how more than 40 partners are adding inference, fine-tuning, RL, agents, and clouds to complete the stack around Beam. It also says reproducible eval recipes are coming soon.
- Hugging Face BlogOfficialPickReflection releases Beam, a 501B open MoE model for coding and agents
AIReflection has released Beam, a 501B-parameter mixture-of-experts model with 23B active parameters per token, under the Apache 2.0 license. It supports a 1M-token context window and targets coding and agentic use, with day-0 support in transformers, SGLang, and vLLM. The post also provides checkpoints in BF16, FP8, and NVFP4, plus inference, SFT, and RL recipes and a Hub credit program for projects built on Beam.
- vLLM BlogOfficialPickvLLM adds day-0 serving for Reflection's Beam 501B MoE model
AIvLLM supports Reflection's Beam, a 501B-parameter MoE with 23B active parameters per token, from day 0 and optimizes it on NVIDIA Blackwell. Agent-driven kernel work cut batch-1 decode from 8.95 ms to 3.30 ms per step on TP8. A DFlash2 drafter with adaptive verification reaches 708 tokens/s on a single request and adds 10.8% throughput at concurrency 128.