MA

Marco Mini Global Base

MarcoGGUF by mradermacherUpdated 2026-06-27
Download on Hugging Face
Parameters17.4B
Context32K
Layers28
Architectureqwen3moe
Smallest quant0.1 GB
Quick answer

Marco Mini Global Base needs about 12.0 GB of memory at Q4_K_M (a 10.6 GB file) with an 8K context window, so it fits on a 12 GB GPU. This repository has 25 quants, from 0.1 GB to 14.3 GB (Q6_K). Longer context needs more: at 32K tokens the KV cache alone adds about 3.8 GB.

17.4B‑parameter multilingual MoE model with 64‑language coverage

Marco Mini Global Base is a 17.4B‑parameter decoder‑only Transformer with sparse MoE layers. It activates 0.86B parameters per token and supports 64 languages. The model was trained on 5.5T tokens across four stages.

Good at

  • Multilingual text generation
  • Sparse MoE efficiency
  • Large context support

Limits

  • No vision support
  • Requires 8k context for memory estimates
  • Apache‑2.0 license

Best for

  • Chatbots needing many languages
  • Research on sparse MoE models
  • Low‑latency inference on GPUs

Who it’s for. Developers and researchers building multilingual applications who need a balance between parameter count and active usage. The model is suitable for environments where GPU memory is limited but a wide language set is required.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
—0.1 GB 1.5 GB
Fits Get file →
IQ1_S3.6 GB 5.1 GB
Fits Get file →
IQ1_M4.0 GB 5.4 GB
Fits Get file →
IQ2_XXS4.7 GB 6.1 GB
Fits Get file →
IQ2_XS5.2 GB 6.6 GB
Fits Get file →
IQ2_S5.2 GB 6.7 GB
Fits Get file →
IQ2_M5.8 GB 7.2 GB
Tight Get file →
Q2_K_S6.0 GB 7.4 GB
Tight Get file →
Q2_K6.4 GB 7.8 GB
Tight Get file →
IQ3_XXS6.7 GB 8.2 GB
Too big Get file →
IQ3_XS7.2 GB 8.6 GB
Too big Get file →
IQ3_S7.6 GB 9.0 GB
Too big Get file →
Q3_K_S7.6 GB 9.0 GB
Too big Get file →
IQ3_M7.7 GB 9.1 GB
Too big Get file →
Q3_K_M8.4 GB 9.8 GB
Too big Get file →
Q3_K_L9.0 GB 10.5 GB
Too big Get file →
IQ4_XS9.3 GB 10.8 GB
Too big Get file →
IQ4_NL9.9 GB 11.3 GB
Too big Get file →
Q4_09.9 GB 11.3 GB
Too big Get file →
Q4_K_S9.9 GB 11.4 GB
Too big Get file →
Q4_K_M10.6 GB 12.0 GB
Too big Get file →
Q4_110.9 GB 12.4 GB
Too big Get file →
Q5_K_S12.0 GB 13.5 GB
Too big Get file →
Q5_K_M12.4 GB 13.8 GB
Too big Get file →
Q6_K14.3 GB 15.7 GB
Too big Get file →

For 8 GB VRAM, use IQ1_S (3.63 GB, 5.1 GB memory) or IQ1_M (4.01 GB, 5.4 GB memory). For 16 GB VRAM, IQ2_M (5.76 GB, 7.2 GB memory) or Q2_K (6.4 GB, 7.8 GB memory) are good choices. Larger VRAM allows higher‑quality quants such as Q3_K_L (9.05 GB, 10.5 GB memory).

What GPU do you need for Marco Mini Global Base?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: the best quant that fits is the main file (0.1 GB file, about 1.5 GB in use).
  • 6 GB VRAM: the best quant that fits is IQ1_M (4.0 GB file, about 5.4 GB in use).
  • 8 GB VRAM: the best quant that fits is Q2_K (6.4 GB file, about 7.8 GB in use).
  • 12 GB VRAM: the best quant that fits is Q4_K_M (10.6 GB file, about 12.0 GB in use).
  • 16 GB VRAM: the best quant that fits is Q6_K (14.3 GB file, about 15.7 GB in use).

Which Marco Mini Global Base quant should you choose?

Q4_K_M is the usual starting point for Marco Mini Global Base: it balances file size and quality well. Here it is a 10.6 GB file.

If memory is tight, the smallest file is the first one in the table at 0.1 GB, with the quality trade-offs that come with fewer bits.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architectureqwen3moe
Parameters17.4B (approximate, mixture of experts)
Layers28
Trained context32,768 tokens
KV cache (fp16)112 KB per token

Source

Quantized bymradermacher
FamilyMarco
Licenseapache-2.0
Downloads (HF)1,456
Last updated2026-06-27

Frequently asked questions about Marco Mini Global Base

How much VRAM does Marco Mini Global Base need?

About 12.0 GB at Q4_K_M with an 8K context window. That is the 10.6 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Marco Mini Global Base on an 8 GB GPU?

Yes, with Q2_K: the 6.4 GB file needs about 7.8 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of Marco Mini Global Base GGUF?

It depends on the quant: file is 0.1 GB, IQ1_S is 3.6 GB, IQ1_M is 4.0 GB, IQ2_XXS is 4.7 GB, IQ2_XS is 5.2 GB, IQ2_S is 5.2 GB and 19 more.

What context length does Marco Mini Global Base support?

The GGUF header lists a trained context length of 32,768 tokens (32K). The KV cache grows with the context you use: about 0.9 GB at 8K and 3.8 GB at 32K.

Where can I download Marco Mini Global Base GGUF?

The files are on Hugging Face in the mradermacher/Marco-Mini-Global-Base-i1-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Marco Mini Global Base use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run Marco Mini Global Base on 8 GB VRAM?

Yes, using IQ1_S or IQ1_M quants keeps VRAM usage under 8 GB and memory usage around 5–5.4 GB.

Does the model support vision tasks?

No, the model is purely text‑based and has no vision capabilities.

What is the activation ratio of the MoE layers?

The model activates 0.86B out of 17.3B total parameters, a 5% activation ratio per token.

Other GGUFs of Marco-Mini-Global-Base

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More Marco models

All Marco →

Other medium models

All medium models →