Reflection Beam 501B
Reflection
Parameters
501.0B total / 23.0B active
MoE: total / active
Architecture
MoE Transformer (per Ripple-65B Preview card comparison table)
Released
—
License
—
About
Comparison-anchor stub record created during the ricursiveAI/Ripple-65B-Preview ingest. The Ripple-65B Preview model card compares its instruct model against this model; parameters taken verbatim from that comparison table (501B total, 23B active, 23.8T pretraining tokens). Full ingestion pending - needs a dedicated HuggingFace-page ingest.
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
SciCode
|
82.29%
|
— |
#3/27
better than 92%
|
|
FORTRESS Benign
safety
|
28.57%
|
— |
#4/5
better than 25%
|
|
StrongREJECT v2 (adaptive)
safety
|
97.15%
|
— |
#4/5
better than 25%
|
|
Terminal Bench 2.1
coding_agent
|
88.36%
|
— |
#18/44
better than 60%
|
|
AutomationBench (Public)
general_agent
|
61.43%
|
— |
#4/11
better than 70%
|
|
GPQA Diamond
stem_reasoning
|
93.00%
|
— |
#26/80
better than 68%
|
|
SWE-bench Verified
coding_agent
|
93.12%
|
— |
#3/60
better than 97%
|
|
DeepSWE 1.1
coding_agent
|
56.12%
|
— |
#24/34
better than 30%
|
|
SWE-bench Pro
coding_agent
|
73.69%
|
— |
#6/77
better than 93%
|
|
SWE-bench Pro v2.0 Hard
coding_agent
|
94.63%
|
— |
#2/5
better than 75%
|
|
AIME 2025
stem_reasoning
|
99.43%
|
— |
#2/24
better than 96%
|
|
AIME 26
stem_reasoning
|
98.09%
|
— |
#7/45
better than 86%
|
|
AA-LCR
long_context
|
93.10%
|
— |
#6/42
better than 88%
|
|
HMMT Feb 26
stem_reasoning
|
95.08%
|
— |
#11/44
better than 74%
|
|
TauBench V3 Banking
general_agent
|
93.35%
|
— |
#2/17
better than 94%
|
|
MMLU
knowledge
|
99.39%
|
— |
#2/16
better than 93%
|
|
AIME 2024
stem_reasoning
|
79.31%
|
— |
#4/7
better than 50%
|
|
Terminal-Bench 2.1 (Terminus-2)
coding_agent
|
73.45%
|
— |
#18/44
better than 58%
|
|
MCP-Atlas
general_agent
|
89.03%
|
— |
#8/54
better than 87%
|
|
MMLU-Pro
knowledge
|
53.23%
|
— |
#33/47
better than 28%
|
|
IFBench
instruction_following
|
92.52%
|
— |
#7/42
better than 85%
|
What is a good SciCode score? What is a good FORTRESS Benign score? What is a good StrongREJECT v2 (adaptive) score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 23.0B are active per token
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 1002.0 GB | 77.8 GB | 1129.8 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 501.0 GB | 38.9 GB | 564.9 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 250.5 GB | 19.4 GB | 282.5 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 250.5 GB | 19.4 GB | 282.5 GB | Which GPUs can run it? Calculate on your GPU |
Only datacenter GPUs — FP16 needs 1129.8 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.
No comments yet
Be the first to share your thoughts!