Kolibri 1 — Aleph Alpha's sovereign open-weight MoE: 78B total with about 3.5B active, trained from scratch on a German- and English-focused corpus. Four in five layers use a 513-token sliding window, so its KV cache stays small at long context.
chatreasoningmultilingualcodingagentictool_use
78.1B
Parameters (3.46B active)
256K
Context length
6
Benchmarks
16
Quantizations
23K
HF downloads
Architecture
MoE
Released
2026-10-03
Layers
50
KV Heads
4
Head Dim
128
Family
other
Quantization Options
Context length:
Quant
Bits
VRAM @ 16K
Quality
IQ2_M
2.93
29.4 GB
29.1 + 0.3 KV
low
Q2_K
3.16
31.6 GB
31.3 + 0.3 KV
low
IQ3_XXS
3.25
32.5 GB
32.2 + 0.3 KV
low
IQ3_XS
3.5
34.9 GB
34.7 + 0.3 KV
low
Q3_K_S
3.64
36.3 GB
36.0 + 0.3 KV
low
IQ3_M
3.76
37.5 GB
37.2 + 0.3 KV
low
Q3_K_M
4
39.8 GB
39.5 + 0.3 KV
low
Q3_K_L
4.3
42.7 GB
42.5 + 0.3 KV
moderate
IQ4_XS
4.46
44.3 GB
44.0 + 0.3 KV
moderate
Q4_K_S
4.67
46.4 GB
46.1 + 0.3 KV
moderate
Q4_K_M
4.89
48.5 GB
48.2 + 0.3 KV
good
Q5_K_S
5.57
55.1 GB
54.9 + 0.3 KV
good
Q5_K_M
5.7
56.4 GB
56.1 + 0.3 KV
good
Q6_K
6.56
64.8 GB
64.5 + 0.3 KV
excellent
Q8_0
8.5
83.7 GB
83.5 + 0.3 KV
lossless
FP16
16
157.0 GB
156.7 + 0.3 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.