Reflection AI unveils Beam, a 501B open-weight model. Weights land later this month under Apache 2.0
Reflection AI has unveiled Beam, its first open-weight model: a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agent work (Reflection, Oct 5, 2026). An early-access waitlist is open now. Reflection says the weights, technical report and model card come later this month under an Apache 2.0 license.
Kingfish737 at the table: a big U.S. open-weight model aimed right at the Chinese open models. The real test comes when the weights ship and people can run it themselves.
The specs Reflection published
- 501B total / 23B active parameters, sparse MoE. Text-only.
- Pretrained on 23.8 trillion tokens from the web, public sources and licensed datasets. Midtraining extends context to 1 million tokens.
- Reinforcement learning ran on 10.5K NVIDIA GB300 GPUs for four weeks and produced more than 100 million rollouts.
- A reasoning effort setting lets you trade shorter answers for longer, harder thinking.
What Reflection claims (not independently verified)
- On advanced reasoning benchmarks, Reflection says Beam scores comparable to Z.ai’s GLM-5.2 while using 3–4× less inference compute.
- Reflection’s own reported scores include 80.9 on SWE-bench Verified, 80.1 on Terminal Bench v2.1 and 90.5 on GPQA Diamond.
- Reflection says Beam is competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding and agent tasks, and admits Kimi K3 is still ahead on raw capability. Its pitch is efficiency.
TechCrunch notes the performance claims haven’t been independently verified. It also reports that Reflection is a two-year-old Brooklyn startup founded by two former Google DeepMind researchers and has raised about $4.7 billion (per PitchBook).
When you can get it
- Now: a waitlist for a select group of early users while Beam finishes red-teaming and evaluations.
- Later this month: weights under Apache 2.0, plus the technical report, model card, docs and tools for running, evaluating and fine-tuning it. Reflection says distribution partners and open-source integrations will be ready at launch.
Neon take: 23B active parameters is the number that matters. If the efficiency claims hold up, this is a model a lot of teams can actually afford to run.
Filed Oct 5, 2026 ~4:28 PM CT — Afternoon pass. Sources: Reflection — Introducing Beam: Reflection’s 501B open-weight model (Oct 5, 2026) · TechCrunch — Reflection debuts Beam (Rebecca Bellan, Oct 5, 2026). Benchmark and compute-efficiency figures are Reflection’s own. No exact release date for the weights beyond “later this month,” and no pricing was announced.