Aleph Alpha, Germany · Released October 2, 2026New
Kolibri-1 has 78.1B parameters (3.46B active per token). In BF16 its weights alone take 145 GiB; quantized to 4-bit, about 44.0 GiB. Each 32K-token request adds 0.66 GiB of KV cache. Cheapest way to run it today: 3 × GeForce RTX 3090 (24 GB) at $0.429 per hour.
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
How many GPUs to run Kolibri-1
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
FP8 (native)
BF16 / FP16
FP8 / INT8
4-bit (GGUF Q4_K_M)
GeForce RTX 4090 24 GB
4
7
4
3
GeForce RTX 5090 32 GB
3
6
3
2
RTX PRO 6000 Blackwell 96 GB
1
2
1
1
Mac M5 Max 128 GB 96 GB usable
1
2
1
1
NVIDIA H100 80 GB
2
3
2
1
NVIDIA H200 141 GB
1
2
1
1
NVIDIA B200 (HGX) 180 GB
1
1
1
1
Cheapest way to run Kolibri-1 today
For each GPU: the fewest cards that fit Kolibri-1 with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-10).
On-demand prices from RunPod, Vast.ai, Verda and Azure, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
KV cache: how memory grows with context
Architecture: MoE 384 experts · 10 full-attention layers + 40 sliding-window layers (513 tokens), GQA 4 KV × 128. Its KV cache is 4.7× smaller than if every layer used full attention, because most layers only keep a short window of recent tokens. Only 3.46B parameters are active per token, but all 78B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
0.20 GiB
0.10 GiB
32K tokens
0.66 GiB
0.33 GiB
128K tokens
2.54 GiB
1.27 GiB
256K tokens
5.04 GiB
2.52 GiB
Kolibri-1 VRAM FAQ
How much VRAM does Kolibri-1 need?
In BF16 the weights alone take 145 GiB (156 GB). In FP8 that is 73.3 GiB, and about 44.0 GiB with 4-bit quantization (Q4_K_M); the official FP8 checkpoint is 73.4 GiB. Each request then adds KV cache: 0.20 GiB at 8K tokens and 0.66 GiB at 32K (BF16 cache).
Can Kolibri-1 run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 3 RTX 4090s. In 4-bit it fits on a single RTX PRO 6000 Blackwell (96 GB).
What is the cheapest way to run Kolibri-1?
On 2026-10-10, the cheapest on-demand setup is 3 × GeForce RTX 3090 (24 GB) with 4-bit (GGUF Q4_K_M) weights on Vast.ai, at $0.429 per hour (about $313.17 per month). Next: 1 × RTX 6000 Ada (48 GB) with 4-bit (GGUF Q4_K_M) weights on Vast.ai, at $0.494 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does Kolibri-1 need?
With one request and an 8K-token context: 3 in BF16, 2 in FP8 and 1 in 4-bit (vLLM, 90% of memory usable). Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Kolibri-1 VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/kolibri-1.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.
Aleph Alpha, Germany · Released October 2, 2026