gigaam v3 needs about 0.8 GB of memory at Q8_0 (a 0.2 GB file) with an 8K context window, so it fits on a 4 GB GPU. This repository has 12 quants, from 0.2 GB (Q4_K) to 0.5 GB (F16).
220‑M‑parameter Russian ASR model in GGUF format
gigaam v3 is a Conformer‑based foundation model for Russian speech recognition. It contains 0.221 billion parameters and is available in several GGUF quantisations. The model is licensed under MIT and can be used with llama.cpp or other GGML backends.
Good at
- Russian ASR
- End‑to‑end CTC and RNN‑T decoding
- Punctuation and casing in transcripts
Limits
- No vision support
- Requires 16 kHz mono audio
- No external training data provided
Best for
- Transcribing Russian phone calls
- Speech‑to‑text for call‑center analytics
- Generating punctuated transcripts for Russian media
Who it’s for. Developers building Russian speech‑to‑text pipelines or integrating ASR into applications. The model is suitable for users who need a lightweight, open‑source solution with a permissive MIT license.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K | 0.2 GB | 0.7 GB | Fits | Get file → | |
| Q4_K | 0.2 GB | 0.7 GB | Fits | Get file → | |
| Q4_K | 0.2 GB | 0.7 GB | Fits | Get file → | |
| Q4_K | 0.2 GB | 0.7 GB | Fits | Get file → | |
| Q8_0 | 0.2 GB | 0.8 GB | Fits | Get file → | |
| Q8_0 | 0.2 GB | 0.8 GB | Fits | Get file → | |
| Q8_0 | 0.2 GB | 0.8 GB | Fits | Get file → | |
| Q8_0 | 0.2 GB | 0.8 GB | Fits | Get file → | |
| F16 | 0.4 GB | 1.0 GB | Fits | Get file → | |
| F16 | 0.4 GB | 1.0 GB | Fits | Get file → | |
| F16 | 0.5 GB | 1.0 GB | Fits | Get file → | |
| F16 | 0.5 GB | 1.0 GB | Fits | Get file → |
For 8 GB VRAM, use the Q8_0 quantisation (0.25 GB file, ~0.8 GB memory at 8k context). If VRAM is limited to 4 GB, choose Q4_K (0.15 GB file, ~0.7 GB memory). For highest precision, F16 (0.45 GB file, ~1 GB memory) is available but requires more memory.
What GPU do you need for gigaam v3?
These figures add the file size, an fp16 KV cache at an 8K context window 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 F16 (0.5 GB file, about 1.0 GB in use).
Which gigaam v3 quant should you choose?
Q8_0 (0.2 GB) is the middle option in this repository. Close to lossless. It costs roughly twice the memory of a 4-bit file, so it suits cases where you have plenty of VRAM.
If memory is tight, the smallest file is Q4_K at 0.2 GB, with the quality trade-offs that come with fewer bits. The usual sweet spot between file size and quality, and the quant most people start with.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about gigaam v3
How much VRAM does gigaam v3 need?
About 0.8 GB at Q8_0 with an 8K context window. That is the 0.2 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 gigaam v3 on an 8 GB GPU?
Yes, with F16: the 0.5 GB file needs about 1.0 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of gigaam v3 GGUF?
It depends on the quant: Q4_K is 0.2 GB, Q4_K is 0.2 GB, Q4_K is 0.2 GB, Q4_K is 0.2 GB, Q8_0 is 0.2 GB, Q8_0 is 0.2 GB and 6 more.
Where can I download gigaam v3 GGUF?
The files are on Hugging Face in the cstr/gigaam-v3-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 gigaam v3 use?
Hugging Face lists the license as mit. 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 gigaam v3 on 4 GB VRAM?
Yes, the Q4_K quantisation uses 0.15 GB file size and requires about 0.7 GB memory at 8k context, fitting within 4 GB VRAM.
Does the model support English speech?
No, the model is trained exclusively on Russian speech data and is intended for Russian ASR tasks.
What is the practical window for long audio?
The model has a practical window of about 25 seconds for full attention; longer inputs are sliced by the CLI's VAD/chunking.