estimate, except measured file sizes

Can I run Kolibri-1 locally?

78.1B total, 3.46B active. Parameter count is an estimate.

Licence:

Apache-2.0 source

Context:

262,144 tokens

Modality:

text (German and English)

Released:

3 Oct 2026 source

Aleph Alpha GmbH's open MoE for German and English, Apache-2.0. The card lists 78,103,074,560 parameters (headline 78B; 78.1B here is that count to one decimal) and 3,457,573,120 active per token (3.46B). Fewer active parameters means less compute per token than a dense model of this total size. That is not a measured tokens-per-second figure: no independent measurement is stored, and speed is not estimated for MoE. 384 experts per layer, 6 active per token, plus a shared expert. llama.cpp b11709 (17 Oct 2026) describes 384 routed experts plus one shared and adds support: https://github.com/ggml-org/llama.cpp/releases/tag/b11709. 50 layers: 40 sliding-window (window 513) and 10 full attention. 48 query heads, 4 KV heads, head dim 128. Config max_position_embeddings is 262,144. The card also validates 1,048,576 tokens and recommends at most 262,144; this catalog uses 262,144. KV in the table is this site's FP16 formula on those 10 full layers and the 513-token window. The card evaluates with an FP8 KV cache, which this formula does not use. bartowski GGUF sizes are decimal GB, shards summed, checked 18 Oct 2026: IQ2_XXS 22.77, Q2_K 29.92, IQ3_XXS 33.65, Q3_K_M 37.62, IQ4_XS 43.63, Q4_K_S 46.40, Q4_K_M 49.76, Q5_K_M 60.30, Q6_K 69.00, Q8_0 83.14, BF16 156.31. The table rounds those files to one decimal. IQ4_XS at 43.63 GB, with the KV and 0.8 GB overhead estimates, fits this catalog's scored 64 GB-class unified-memory machines. That fit is the formula, not a test. Q4_K_M at 49.76 GB fits the 64 GB Macs on the same formula (54 GB usable after the macOS reserve). The 64 GB Strix Halo row is scored at 48 GB dedicated graphics, enough for IQ4_XS and not for Q4_K_M. The card's own hardware line is about 78 GB of FP8 weights on data-center GPUs; that is a card claim, not a measurement here. The card's release date is 3 Oct 2026. The Hugging Face repo was created 2 Oct 2026. No official Ollama tag on 18 Oct 2026.

Parameter source · Architecture source · Model page

speed: not estimated for MoE yet

Memory per quant

KV is at 8,192 tokens. Overhead is 0.8 GB. Both are estimates. A measured file size is labelled. The last column is the cheapest machine that fits the total fully.

Cheapest rig that fits

Q4, full: Mac mini M4 Pro 64GB, €1,890 (estimate)

Q4, slow, partly in system RAM: Budget PC with RTX 3060 12GB, slow, partly in system RAM, €1,357 (estimate)

IQ2_XXS (smallest file), full: Mac mini M6 32GB, €1,530 (estimate)

IQ2_XXS (smallest file), slow, partly in system RAM: Budget PC with RTX 3060 12GB, slow, partly in system RAM, €1,357 (estimate)

Prices moved: see Building a local LLM rig in the RAMageddon.

On a specific GPU

• Kolibri-1 on RTX 3060 12GB
• Kolibri-1 on RTX 4060 Ti 16GB
• Kolibri-1 on RTX 3090
• Kolibri-1 on RTX 4090
• Kolibri-1 on RTX 5080
• Kolibri-1 on RTX 5090
• Kolibri-1 on RX 7900 XTX
• Kolibri-1 on Dual RTX 3090
• Kolibri-1 on Dual RTX 4090