Beam 1.0 is Reflection AI's first model: a 501B MoE with 23B active, trained from scratch, English-first, with a 1M context and sliding-window attention on most layers. Reflection pitches it on token-efficient reasoning, claiming similar answers with about a third of the thinking tokens.
chatcodingreasoningmathagentictool_use
501.27B
Parameters (23.4B active)
1024K
Context length
8
Benchmarks
17
Quantizations
10K
HF downloads
Architecture
MoE
Released
2026-10-16
Layers
52
KV Heads
8
Head Dim
128
Family
other
Quantization Options
Context length:
Quant
Bits
VRAM @ 16K
Quality
IQ2_XXS
2.38
150.2 GB
149.6 + 0.6 KV
low
IQ2_M
2.93
184.6 GB
184.1 + 0.6 KV
low
Q2_K
3.16
199.0 GB
198.5 + 0.6 KV
low
IQ3_XXS
3.25
204.7 GB
204.1 + 0.6 KV
low
IQ3_XS
3.5
220.3 GB
219.8 + 0.6 KV
low
Q3_K_S
3.64
229.1 GB
228.6 + 0.6 KV
low
IQ3_M
3.76
236.6 GB
236.1 + 0.6 KV
low
Q3_K_M
4
251.7 GB
251.1 + 0.6 KV
low
Q3_K_L
4.3
270.5 GB
269.9 + 0.6 KV
moderate
IQ4_XS
4.46
280.5 GB
279.9 + 0.6 KV
moderate
Q4_K_S
4.67
293.7 GB
293.1 + 0.6 KV
moderate
Q4_K_M
4.89
307.4 GB
306.9 + 0.6 KV
good
Q5_K_S
5.57
350.0 GB
349.5 + 0.6 KV
good
Q5_K_M
5.7
358.2 GB
357.6 + 0.6 KV
good
Q6_K
6.56
412.1 GB
411.5 + 0.6 KV
excellent
Q8_0
8.5
533.6 GB
533.1 + 0.6 KV
lossless
FP16
16
1003.6 GB
1003.0 + 0.6 KV
lossless
Select your GPU above to see speed estimates and compatibility for each quantization.
Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.