Dataset updated 2026-10-24

Kolibri 1

Kolibri 1 loads in 49.8 GB at Q4_K_M and decodes at the speed of its 3.5B active path, comfortably from 96 GB of memory, Macs included, starting at the MacBook Pro M5 Max 128GB. Below is the memory math we ran for it, the smallest card that fits, and the fastest machine we track.

Parameters
78.1B(3.5B active)
Format
Q4_K_M
Min memory
96 GB
Best for
Reasoning, German, Long context
Memory math

How much memory does Kolibri 1 need?

Kolibri 1 is a 78.1B mixture-of-experts model (3.5B active per token) built for reasoning, german, long context. The weights alone take 49.8 GB at Q4_K_M; add context and the OS, and it belongs in the 96 GB memory class. Every number below comes from the ModelFit engine: speeds are estimates, labeled est., because we do not benchmark hardware ourselves.

Weights (Q4_K_M)
49.8 GB
+ KV cache (16k)
~0.3 GB
Total at 16k
~50.1 GB
Comfortable from
96 GB

KV = fp16 estimate (q8_0 cache roughly halves it). "Comfortable" = weights + KV within the engine's tiered budget (~70-85% of memory).

Hardware snapshot

Which machines run Kolibri 1?

Smallest GPU
NVIDIA RTX PRO 6000 Blackwellest. 282 tok/s · 96 GB VRAMtypical error 15%
Fastest
NVIDIA RTX PRO 6000 Blackwellest. 282 tok/stypical error 15%
Runs on a Mac?
MacBook Pro M5 Max 128GBest. 121 tok/stypical error 10%

Verdicts come from the same engine that powers our fit checker: memory capacity first, bandwidth for speed. How we compute these numbers · What typical error means · Hugging Face model card

Go deeper

  • Also tracked: Kolibri 1 (IQ4_XS)(IQ4_XS, 43.6 GB, min 64 GB)

Frequently asked questions

How much memory does Kolibri 1 need?

49.8 GB for the Q4_K_M weights, plus ~0.3 GB of KV-cache at 16k context, about 50.1 GB total. Comfortable from 96 GB of VRAM or unified memory.

Does Kolibri 1 run on a Mac?

Yes, from the MacBook Pro M5 Max 128GB (est. 121 tok/s, typical error 10%). Unified memory means the RAM budget is the only limit.

What is the smallest GPU for Kolibri 1?

The NVIDIA RTX PRO 6000 Blackwell (96 GB VRAM) is the smallest tracked card that runs Kolibri 1 comfortably, est. 282 tok/s, typical error 15%.

Cite this page

ModelFit: Kolibri 1, specs, memory math and hardware verdicts.
https://modelfit.io/models/kolibri-1/ (dataset updated 2026-10-24, CC BY 4.0).