News · SI
Reflection has published the weights for Beam, a 501B-parameter mixture-of-experts model with 23B active and a 1M-token context, under the Apache 2.0 licence. It ships with an open-source agent harness, day-one support in vLLM and SGLang, and availability on more than 20 clouds and inference providers.
Reflection has released the weights for Beam, its first open-weight model, on Hugging Face under the Apache 2.0 licence[^1]. Beam is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning and long-running agent work, with a 1 million token context window[^1][^2].
01 / 05
Reflection pretrained Beam on 23.8 trillion tokens, then ran more than 100 million reinforcement learning rollouts on 10,500 NVIDIA GB300 GPUs over four weeks[^3]. Training and grading used about 1.3 billion sandboxes, and Reflection says capabilities were still climbing as it added compute[^3]. The model is text only and has 52 layers with 256 experts, eight of which run for each token[^2].
02 / 05
Reflection reports 80.9 on SWE-bench Verified, 80.1 on Terminal Bench 2.1 and 90.5 on GPQA Diamond[^3]. Its own comparison is candid: Kimi K3 scores higher on most tests, and DeepSeek V4.1 Flash leads it on DeepSWE v1.1 (74.2 against 44.4) and Terminal Bench 2.1 (90.6 against 80.1)[^3]. Reflection's pitch is efficiency instead, saying Beam matches GLM-5.2 on advanced reasoning tests with three to four times less inference compute[^3].
03 / 05
Beam ships with Mirror CLI, an open-source coding harness Reflection trained the model alongside[^1]. Mirror runs from a terminal, lets you approve tool calls, resume conversations and run parallel agents in separate worktrees, and supports skills, MCP connections and Python extensions[^1]. Beam has also been tested in Pi, OpenCode and Hermes[^1].
Weights come in BF16, FP8 and NVFP4 with a speculative decoding drafter, and Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo support it from day one[^1]. Hugging Face puts the BF16 checkpoint at about 1 TB and the NVFP4 one at about 280 GB, and suggests four B200 GPUs to leave room for long contexts[^2].
04 / 05
Reflection's own API at platform.reflection.ai is in public beta and free to start[^1]. Beam is also on AWS, Google Cloud, Microsoft Foundry, Cloudflare, Databricks, Fireworks AI, Together AI, OpenRouter, Ollama, LM Studio, Vercel and others, and Dell and HPE deliver it for on-premises and air-gapped sites[^1].
05 / 05
Beam is not the strongest open model on coding benchmarks, and Reflection says so. Its case is the licence and the package: Apache 2.0 weights from a US lab, a matching open-source harness, and fine-tuning and RL recipes ready to use. If your team needs a model it can legally modify, host behind its own firewall and post-train on internal code, Beam is worth a trial on a hosted endpoint before you commit GPUs.
FAQ
Beam is Reflection's first open-weight model. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning and agent work. It takes text only and supports a 1 million token context window.
Reflection announced Beam on 5 October 2026 and released the weights on Hugging Face on 16 October 2026 (US time). They are under the Apache 2.0 licence, which allows commercial use and modification.
All 256 experts stay in memory, so you need room for the full 501B model. Hugging Face says the BF16 checkpoint is about 1 TB and the NVFP4 checkpoint about 280 GB. It suggests four B200 GPUs for comfortable serving, and shows the NVFP4 build on four B300s and the BF16 build across two 8x H200 nodes with vLLM.
Reflection's own API at platform.reflection.ai is in public beta and free to start. Beam is also available through partners including Amazon Web Services, Google Cloud, Microsoft Foundry, Cloudflare, Databricks, Fireworks AI, Together AI, OpenRouter, Ollama, LM Studio and Vercel, and Dell and HPE offer it for on-premises and air-gapped setups.
Mirror is Reflection's open-source coding agent harness, which Beam was trained alongside. It runs from a terminal, supports resumable conversations, parallel agents in separate worktrees, skills, MCP connections and Python extensions, and uses a Rust agent core with a Python frontend. Beam has also been tested with Pi, OpenCode and Hermes.
Reflection reports 80.9 on SWE-bench Verified and 80.1 on Terminal Bench 2.1. Its own table also shows Kimi K3 and DeepSeek V4.1 Flash ahead on DeepSWE v1.1 and Terminal Bench 2.1, and Reflection says Beam's advantage is efficiency rather than raw capability. These are Reflection's figures, so test it on your own work.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
