Apparently October is open-weights month. On the same day Mistral rolled out its trillion-parameter heavyweight, Reflection showed up with Beam, a model built on a simple idea: you don't need the whole orchestra playing to get through every song.

Big Roster, Small Starting Lineup

Reflection introduced Beam on October 6 as a sparse Mixture-of-Experts model with 501 billion total parameters, of which only 23 billion are active per token. It was pretrained on 23.8 trillion tokens from web and proprietary licensed data using 6,144 NVIDIA GB300 GPUs over four weeks, then put through heavy reinforcement learning with more than 100 million rollouts on 10.5K GPUs over another four weeks. The context window is 256K tokens and can be extended to 1 million.

Beam is released under Apache 2.0, one of the most permissive licenses out there. For now it's in early access, with full weights, a technical report and other artifacts coming later in October. Reflection says the model was tuned for coding, reasoning and agentic work, and that it performs close to models like Qwen 3.8-Max while using "3–4× less inference compute" on hard reasoning tasks.

The Dial Is the Feature

The spec sheet will grab the headlines, but the clever bit is the controllable reasoning effort. Users can tell Beam how hard to think, matching compute to the task. That sounds like a minor setting until you get the bill for a reasoning model that spent four thousand tokens deciding whether "Hi" needed a reply. Paying for effort only where you need it is how AI features stay affordable once they're in production.

Apache 2.0 matters too. Plenty of "open" models come with licenses that need a lawyer and a cup of strong coffee to interpret, and Apache 2.0 isn't one of them. The usual caveats apply: these are the company's own benchmarks, and the weights aren't out yet, so give it a few weeks of independent testing before calling a winner.

Compute is expensive, and models that know when to take it easy are going to win a lot of budget meetings.

If your AI feature costs are climbing faster than its usefulness, let's look at whether a leaner model or smarter routing could do the same job for less.

Source: Reflection AI