Reflection’s new open-weight model, Beam, is now available through Baseten’s Model APIs and Dedicated Inference. The launch gives developers a model pitched for coding and agentic work, with adjustable reasoning effort and a one-million-token context window.
Baseten Watch analysis
What happened
Baseten says Beam has 501 billion total parameters, with 23 billion active per token. It is a mixture-of-experts model that accepts text and produces text, mainly in English. Baseten says it works with the runtimes, frameworks, coding harnesses and tools developers already use.
Reflection trained Beam with reinforcement learning, using longer rollouts so it could practise multi-step tasks, use tools and adjust its approach based on feedback. Users can set reasoning effort: lower settings favour shorter responses, while higher settings allow more reasoning for demanding tasks. Read Baseten’s Beam announcement.
Why it matters
A large context window and support for existing developer tools make Beam a substantial addition to Baseten’s hosted model options, particularly for teams building coding and agentic systems. The adjustable effort setting also offers a practical trade-off: spend fewer tokens on simpler work, or allow more reasoning when the task calls for it.
Baseten’s announcement says Reflection’s evaluations show strong results across coding, agentic, reasoning and STEM benchmarks, but the supplied announcement does not give benchmark figures or testing conditions. Treat that as Reflection’s claim, not a settled comparison with other models.
Our read
There is useful substance here beyond a new name in a model menu: developers can try Beam through Baseten’s APIs or dedicated inference, and tune how much reasoning it spends. That makes it worth a look for teams already using the platform. The next question is whether its performance and token savings hold up on real workloads, where benchmark tables have to meet actual code.
What to watch
- How Beam performs on representative coding and agentic workloads.
- Whether lower reasoning settings reduce token use without harming results.
- When Reflection’s planned multilingual and multimodal capabilities arrive.
Discussion spark: For coding and agentic tasks, would you prioritise a model’s benchmark scores or the ability to control reasoning effort and token use?
Sources and evidence
- Source update (16 October 2026, 13:51 UTC)
not affiliated with or endorsed by Baseten