Beam
- Oct 5, 2026
- Oct 5, 2026
Beam is a model from Reflection AI, released Oct 5, 2026. What it could let Clay do, as a hypothesis: could let Clay run enrichment agents on cheaper open weights, so agents switch on inside the plan the buyer already pays for. Every number is vendor-reported; no independent benchmark found yet.
Beam: a model from Reflection AI
Beam is a 501B-parameter open-weight mixture-of-experts model, 23B active, with a 1M-token context window, built for coding, reasoning, tool use and agentic workloads.
- Model
- Vendor-reported only
- Weights under an Apache 2.0 license later this month, with technical report, model card and tools for running, evaluating and fine-tuning.
What Beam could do for Clay on “Agents included”
Could let Clay run enrichment agents on cheaper open weights, so agents switch on inside the plan the buyer already pays for.
- Clay stands at high, on medium evidence.
- Whether AI agents switch on inside the current contract, or arrive as a second line on the invoice.
- Vendor-reported only
Beam: four numbers, all vendor-reported
80.9
- Not stated
- Vendor-reported
3-4x less inference compute
- Z .ai’s GLM-5.2
- Vendor-reported
3–4x higher inference efficiency
- GLM 5.2
- Vendor-reported
over 4x higher efficiency
- leading Western open models in its class
- Vendor-reported
Beam’s evidence: two citations.
Reflection AI, relayed by @kimmonismus: SWE-bench Verified
Reflection AI, relayed by WallStEngine: advanced reasoning benchmarks