Reflection Beam is Reflection AI's first open-weight model, a 501-billion-parameter sparse mixture-of-experts built for coding and agentic software work where teams want frontier reasoning at lower inference cost. It was introduced on October 5, 2026.
Core Features
- 501B total parameters with only 23B active per token, a sparse MoE design that keeps compute per request low.
- Tuned for coding and agentic workloads, including long-horizon software tasks.
- Reflection reports it matches Z.ai's GLM-5.2 on reasoning while using 3 to 4 times less inference compute.
- Apache 2.0 weights planned for later in October 2026, permitting commercial use.
- Targets agent harnesses that run many tool calls and need consistent, cheap decisions.
Use Cases
- Agentic coding platforms that orchestrate long task sequences.
- Self-hosted deployments where open weights avoid per-call API fees.
- Enterprises benchmarking cost per solved issue rather than raw benchmark scores.
Pricing
Beam is positioned as an open-weight release; Reflection AI plans to publish Apache 2.0 weights in October 2026. Until then it is available as a research preview, with no public per-token price yet — the value case is lower inference cost versus dense frontier models.
Pros & Cons
Pros: huge capacity with low active params; open-weight friendly; strong reasoning per compute. Cons: weights not yet public; reasoning still trails the very top dense models on some tasks.
Our Take
Best for teams that run many coding agents and care about cost per task. The trade-off is waiting for the open weights and accepting slightly lower peak reasoning than the largest dense models.
See more AI coding tools on aifreetool.site.
