When the best AI coding model becomes free to run yourself
Reflection AI's new open model Beam matches frontier AI performance on coding tasks and will soon be free for any team to host on its own servers.
- 501B
- total model parameters
- 23B
- parameters active per task (why it runs efficiently)
- 80.9
- score on SWE-Bench Verified (coding benchmark)
- Apache 2.0
- free commercial license, weights coming later in October
Most powerful AI models for coding work the same way: you pay every time you use them. You pay per token — a small chunk of text, roughly three-quarters of a word — and the best models cost between $10 and $30 for every million tokens. For a team writing a lot of code with AI help, those bills can grow very large very fast.
On October 5, a startup called Reflection AI launched a model named Beam. Beam was built specifically for coding and for running automated AI tasks — the kind where the AI takes a series of steps on its own to get something done. It scored 80.9 on SWE-Bench Verified, a standard test that measures how well an AI can fix real software bugs, placing it among the best open models in the world. The weights — the actual files that make the model work — will be released free under an Apache 2.0 license later this month.
The reason Beam can be so efficient is how it is built. It has 501 billion parameters — the internal numbers that control what the model knows and how it thinks — but only 23 billion are active for any given task. This design, called a Mixture-of-Experts, works like a large team where only the right specialists show up for each job. Reflection says this means Beam uses three to four times less computing power than rival open models of similar quality.
For any team building software, running AI agents, or automating repetitive work, this is a meaningful shift. Today, using top-tier AI for those tasks means paying a per-token bill to OpenAI or Anthropic that grows with every action the AI takes. With a model like Beam running on your own servers, the cost becomes a flat, predictable infrastructure expense — and your data never has to leave your systems. The weights are not yet publicly available, but once they land, the economics of AI-assisted work change for any team willing to self-host.