A 501B open model you can host yourself: 23B active parameters per token and a KV cache about one-sixth the naive size
Synopsis
Reflection, a new US lab, released Beam, its first open model: 501B sparse mixture-of-experts weights under Apache 2.0, with about 23B parameters active per token, a 1M-token context, and a focus on coding and agentic work. Hybrid attention keeps only 9 of 52 layers global while 43 hold a fixed 512-token window, so the KV cache grows about 36 KiB per token, roughly one-sixth of the all-global rate. BF16, FP8 and NVFP4 checkpoints ship alongside day-0 transformers, SGLang and vLLM support and a Rust-based harness.
Interpretation
A 501B-parameter open model can now be downloaded, deployed on a team's own hardware or cloud, and modified, rather than reached only through a vendor interface. Teams wanting coding and agentic capability at this scale previously depended on external services; Beam ships its weights and checkpoints under Apache 2.0. It is a sparse mixture-of-experts model in which each token activates about 23B of the 501B parameters, so deployment needs the memory of a 501B model but runs near the speed of a 23B one; the BF16 checkpoint is about 1 TB.
The memory cost of long context drops well below what the same layer count would need with global attention everywhere. If all 52 layers kept every token, the cache would grow by 208 KiB per token; only 9 Beam layers keep the full sequence, and 43 hold a fixed 512-token window and stop growing. In BF16 with the transformers dynamic cache, the cache grows about 36 KiB per token, roughly one-sixth of the naive rate; one sequence needs about 1.2 GiB at 32K tokens, 4.6 GiB at 128K and 36.08 GiB at the full 1M context.
Beyond long context, the release pairs the weights with engineering support for coding and tool calling, so self-hosting is not just a download. It includes a native transformers implementation, day-0 SGLang and vLLM support, and supervised fine-tuning and reinforcement-learning recipes. The package adds a dedicated Rust-based harness, a DFlash2 speculative-decoding drafter, a response template inside the tokenizer, and five reasoning-effort settings.
Perspective
This report covers the model release and its tooling: 501B total parameters, about 23B active per token, a 1M-token context, the Apache 2.0 license, and BF16, FP8 and NVFP4 checkpoints. It is aimed at teams that need to deploy and post-train a large model on their own or cloud hardware; the stated deployment references include four B200s to leave room for long contexts and concurrent requests, or four B300s running the NVFP4 checkpoint on Blackwell.
Benchmark results are provided by the Reflection team across BF16, FP8 and NVFP4 precisions and are not an independent reproduction. The five task trials the release also shares are explicitly called qualitative trials rather than benchmarks; in one low-effort long-context retrieval run the response got all code behavior and all six citations right but did not match the parser format, which the release partly attributes to prompt ambiguity. The multilingual-thinking LoRA evaluation measures language adherence, not the correctness of reasoning.