model
Kolibri-1: how much memory, and what runs it
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.
Kolibri-1 needs about 50.8 GB at Q4_K_M with 8K context (estimate). See which GPUs and Macs fit it, and EU prices. Speed is not estimated for this mixture-of-experts model.
| Quantisation | Estimated memory |
|---|---|
| Q4 | 50.8 GBestimate |
| Q5 | 56.5 GBestimate |
| Q8 | 84.6 GBestimate |
| FP16 | 157.2 GBestimate |
At 8,192 tokens of context. Catalog quant rates are a formula. sourceUrl on that memory stays null.
Parameter count source: huggingface.co/Aleph-Alpha/Kolibri-1. 78.1B total, 3.46B active. Memory math uses the total.
VRAM per quant and the cheapest rig for Kolibri-1
Three setups
Kolibri-1 · Q4 · 8,192 ctx
Estimated memory
50.8 GBestimate
49.8 GB weights · 0.2 GB KV cache · 0.8 GB overhead
01 · Fastest
Mac Studio M5 Ultra 96GB
High bandwidth and enough memory for 70B Q4/Q5. FP16 70B does not fit.
$7,444€6,649
price as of 16 Oct 2026
Belgian Apple StoreBelgian Apple Store price for this config, from the 16 Oct 2026 fact-check.
speed
not estimated for MoE yetspeed: not estimated for MoE yet
needed
50.8 GBfull fit
usable
84.0 GBunified
Why this one
- The model fits entirely in VRAM or unified memory. No CPU offload.
- Memory left for a longer context or a heavier quantisation.
- Unified memory: weights and KV cache share one pool.
- High memory bandwidth, so a higher tokens-per-second estimate.
- One box, little noise, no discrete GPU.
- Highest estimated speed in this shortlist.
Parts
- Mac Studio M5 Ultra, 96 GB, 1.2 TB/s
EU shops
US shops are not linked yet.
02 · A step up
Mac Studio M5 Ultra 256GB
The machine in this list that can still hold 70B in FP16. The euro amount is an estimate.
$13,994€12,500
price as of 16 Oct 2026
price unverifiedEstimate from the 16 Oct 2026 fact-check. Not a checked Apple Store listing. About €12,500–€13,000. This row stores the low end. 256 GB requires the 36-core CPU / 80-core GPU chip.
speed
not estimated for MoE yetspeed: not estimated for MoE yet
needed
50.8 GBfull fit
usable
244.0 GBunified
Why this one
- The model fits entirely in VRAM or unified memory. No CPU offload.
- Memory left for a longer context or a heavier quantisation.
- Unified memory: weights and KV cache share one pool.
- High memory bandwidth, so a higher tokens-per-second estimate.
- One box, little noise, no discrete GPU.
Parts
- Mac Studio M5 Ultra, 256 GB unified memory, 1.2 TB/s
- 256 GB requires the 36-core CPU / 80-core GPU chip
EU shops
US shops are not linked yet.
03 · A step up
Mac Studio M5 Max 128GB
70B up to Q8 has room here. Bandwidth still trails a dual high-end GPU.
$6,717€6,000
price as of 16 Oct 2026
price unverifiedEstimate from the 16 Oct 2026 fact-check. Not a checked Apple Store listing. About €6,000.
speed
not estimated for MoE yetspeed: not estimated for MoE yet
needed
50.8 GBfull fit
usable
116.0 GBunified
Why this one
- The model fits entirely in VRAM or unified memory. No CPU offload.
- Memory left for a longer context or a heavier quantisation.
- Unified memory: weights and KV cache share one pool.
- One box, little noise, no discrete GPU.
Parts
- Mac Studio M5 Max 18-core/40-core GPU, 128 GB, 614 GB/s
EU shops
US shops are not linked yet.
With CPU offload, and therefore slower, this could also work: Dual RTX 4090, 48 GB, speed: not estimated for MoE yet, $8,206€7,330.
Tokens per second are estimated from memory bandwidth, not from a measured benchmark. Prices incl. VAT, excl. shipping. USD = FX estimate, excl. sales tax. A price without a checked listing is marked price unverified. Run llama-bench yourself before you buy.
Want a build tailored to you? Custom report
Get new builds when GPU prices move
End of results
Embed this
[](https://localllmrig.com/can-i-run/kolibri-1/rtx-3060-12gb?utm_source=badge)
Model badge
[](https://localllmrig.com/models/kolibri-1?utm_source=badge)
Common questions
- How much memory does Kolibri-1 need at Q4?
- About 50.8 GB at 8,192 tokens: 49.8 GB weights, 0.2 GB KV cache, 0.8 GB overhead. That is an estimate. sourceUrl stays null.
- Does Kolibri-1 fit on a single 24 GB GPU?
- Not in full. The estimate is 50.8 GB, and a 24 GB card leaves 23.4 GB usable. sourceUrl stays null.
- Are the tokens per second measured?
- speed: not estimated for MoE yet