AH

Ahma-3B

AhmaGGUF by mradermacherUpdated 2024-06-16
Download on Hugging Face
Parameters3.63B
Context2K
Layers26
Architecturellama
Smallest quant2.1 GB
Quick answer

Ahma-3B needs about 6.0 GB of memory at Q4_K_M (a 2.8 GB file) with an 8K context window, so it fits on a 6 GB GPU. This repository has 15 quants, from 2.1 GB (IQ3_S) to 7.3 GB (F16).

Ahma-3B – a 3.63B parameter Finnish language model

Ahma-3B is a 3.63B parameter decoder‑only transformer based on Meta’s Llama v1 architecture. It is pretrained from scratch on Finnish text and supports instruction following out of the box. The model is released under the Apache‑2.0 license.

Good at

  • Finnish text generation
  • Instruction following
  • Fine‑tuning for downstream tasks

Limits

  • No multilingual or code generation
  • Limited to 2048 context tokens
  • Requires at least 5.4 GB VRAM for 8 k context

Best for

  • Local Finnish chatbot
  • Fine‑tuning Finnish NLP tasks
  • Offline text generation

Who it’s for. Developers and researchers working with Finnish language models who need a lightweight, open‑source option. The model is suitable for local inference on consumer GPUs with 8 GB VRAM or more.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ3_S2.1 GB 3.3 GB
Fits Get file →
IQ3_XS2.1 GB 3.3 GB
Fits Get file →
Q2_K2.1 GB 3.3 GB
Fits Get file →
Q3_K_S2.1 GB 3.3 GB
Fits Get file →
IQ4_XS2.2 GB 3.3 GB
Fits Get file →
IQ3_M2.2 GB 3.4 GB
Fits Get file →
Q3_K_M2.3 GB 3.5 GB
Fits Get file →
Q3_K_L2.4 GB 3.6 GB
Fits Get file →
Q4_K_S2.6 GB 3.8 GB
Fits Get file →
Q4_K_M2.8 GB 3.9 GB
Fits Get file →
Q5_K_S2.8 GB 4.0 GB
Fits Get file →
Q5_K_M2.9 GB 4.1 GB
Fits Get file →
Q6_K3.9 GB 5.0 GB
Fits Get file →
Q8_03.9 GB 5.0 GB
Fits Get file →
F167.3 GB 8.4 GB
Too big Get file →

For 8 GB VRAM, use IQ3_S or IQ3_XS (2.15 GB file, ~5.4 GB VRAM). If you have 12 GB VRAM, IQ4_K_M (2.76 GB file, ~6 GB VRAM) offers higher quality. For the best quality with 16 GB VRAM, choose Q5_K_M (2.95 GB file, ~6.2 GB VRAM).

What GPU do you need for Ahma-3B?

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: no quant fits at an 8K context.
  • 6 GB VRAM: the best quant that fits is Q5_K_S (2.8 GB file, about 6.0 GB in use).
  • 8 GB VRAM: the best quant that fits is Q6_K (3.9 GB file, about 7.1 GB in use).
  • 12 GB VRAM: the best quant that fits is F16 (7.3 GB file, about 10.5 GB in use).

Which Ahma-3B quant should you choose?

Q4_K_M is the usual starting point for Ahma-3B: it balances file size and quality well. Here it is a 2.8 GB file.

The Q8_0 file is 3.9 GB, about 1.4 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is IQ3_S at 2.1 GB, with the quality trade-offs that come with fewer bits. A reasonable choice when a 4-bit file will not fit, mostly on larger models. Quality loss is noticeable but often acceptable.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters3.63B
Layers26
Trained context2,048 tokens
KV cache (fp16)325 KB per token

Source

Quantized bymradermacher
FamilyAhma
Licenseapache-2.0
Downloads (HF)2,081
Last updated2024-06-16

Frequently asked questions about Ahma-3B

How much VRAM does Ahma-3B need?

About 6.0 GB at Q4_K_M with an 8K context window. That is the 2.8 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 Ahma-3B on an 8 GB GPU?

Yes, with Q6_K: the 3.9 GB file needs about 7.1 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of Ahma-3B GGUF?

It depends on the quant: IQ3_S is 2.1 GB, IQ3_XS is 2.1 GB, Q2_K is 2.1 GB, Q3_K_S is 2.1 GB, IQ4_XS is 2.2 GB, IQ3_M is 2.2 GB and 9 more.

What context length does Ahma-3B support?

The GGUF header lists a trained context length of 2,048 tokens (2K). The KV cache grows with the context you use: about 2.7 GB at 8K and 0.7 GB at 2K.

Where can I download Ahma-3B GGUF?

The files are on Hugging Face in the mradermacher/Ahma-3B-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 Ahma-3B 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 Ahma-3B on 8 GB VRAM?

Yes, using IQ3_S or IQ3_XS quantizations requires about 5.4 GB VRAM at 8 k context.

Does Ahma-3B support multilingual text?

No, it was trained only on Finnish text and is not suitable for multilingual or code generation.

Other GGUFs of Ahma-3B

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More Ahma models

All Ahma →

Other small models

All small models →