A Lab Just Open-Sourced Its Training Harness (Plus 501B Weights)
Reflection AI dropped Beam this week — 501 billion parameters of open weights under Apache 2.0. That's the headline. But here's what actually caught my eye: they also open-sourced Mirror CLI, their actual training harness. Meta never did that with Llama. I haven't run Beam myself — no independent benchmarks exist yet, so everything below comes from Reflection's release materials and launch…
Reflection AI released Beam, a 501 billion parameter open-weight model, along with its training harness called Mirror CLI this week. This is notable because typically, major labs like Meta only provide the weights and not the training code when releasing open models. Beam is a sparse mixture-of-experts designed for coding, reasoning, and agentic tasks similar to recently released Muse Spark Open 816B model.
The model comes in three weight formats: BF16, FP8, and NVFP4. Mirror CLI itself is the real standout here, as few labs share their actual training harness. This move by Reflection AI puts a lot more power into the hands of researchers and developers who want to fine-tune or pretrain on this 500B+ model. While there are no independent benchmarks yet, comparing Beam to Muse Spark Open gives us a good case study in how training harnesses can impact the open-weight ecosystem.
If you're considering using a permissive-licensed 500B-class model for self-hosting or fine-tuning, or if you're researching training methods, Beam is definitely worth looking at. However, without independent numbers to back up the performance claims, you might want to hold off on using it for production systems just yet.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.