AI Models · Benchmarks
Beam 1.0 vs
DeepSeek V4.1 Flash
A clear side-by-side view of published benchmarks. These are not our own tests.
501BBeam total parameters
23BBeam active parameters
1MBeam context tokens
Benchmark comparison
Higher is better. Figures from Reflection's October 8, 2026 revised launch table.
Beam 1.0DeepSeek V4.1 Flash
Terminal-Bench 2.1
DeepSWE v1.1
GPQA Diamond
| Benchmark | Beam | DeepSeek |
|---|---|---|
| Terminal-Bench 2.1 | 80.1% | 90.6% |
| DeepSWE v1.1 | 44.4% | 74.2% |
| GPQA Diamond | 90.5% | 90.9% |
What do these numbers mean?
In this published comparison, DeepSeek leads on both coding benchmarks. The GPQA Diamond reasoning scores are close. These numbers alone do not establish performance on a real software project.
Sources
- Reflection AI — Introducing Beam (launch table revised Oct 8)
- Reflection AI — Model card (later evaluations)
- Reflection AI — Beam is available (Oct 16)
Note: The later Beam model card reports some different results (e.g., BF16 DeepSWE 43.3% and GPQA Diamond 90.6%). This page preserves the October 8 comparison used for the linked X post. Model checkpoints and evaluation setups can differ.
Enjoyed this research?Leave a like — no account needed.
Loading likes…