Granite Speech 4.1 2B NAR needs about 2.2 GB of memory at Q4_K_M (a 1.6 GB file) with an 8K context window, so it fits on a 4 GB GPU. This repository has 6 quants, from 1.6 GB (Q4_K_M) to 4.5 GB (F16).
Granite Speech 4.1 2B NAR – fast non‑autoregressive speech recognition
Granite Speech 4.1 2B NAR is a 2.3 B parameter model that performs ASR in a single forward pass. It uses a Conformer encoder and a bidirectional LLM editor to edit a CTC hypothesis. The model supports English, French, German, Spanish, and Portuguese.
Good at
- Fast inference with single‑pass decoding
- Competitive WER on LibriSpeech
- Low latency for real‑time ASR
Limits
- No timestamps or speaker labels
- No vision or translation support
- Requires flash_attention_2 and transformers>=5.5.3
Best for
- Real‑time transcription of short audio clips
- Low‑latency voice assistants
- Batch processing of large audio datasets
Who it’s for. Developers building latency‑sensitive speech applications that need multilingual transcription without the overhead of autoregressive decoding. The model is suitable for edge devices and cloud services where inference speed is critical.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 1.6 GB | 2.2 GB | Fits | Get file → | |
| Q5_K_M | 1.8 GB | 2.5 GB | Fits | Get file → | |
| Q6_K | 2.0 GB | 2.7 GB | Fits | Get file → | |
| Q8_0 | 2.5 GB | 3.3 GB | Fits | Get file → | |
| BF16 | 4.5 GB | 5.6 GB | Fits | Get file → | |
| F16 | 4.5 GB | 5.6 GB | Fits | Get file → |
For 8 GB VRAM, Q4_K_M (1.56 GB) or Q5_K_M (1.78 GB) provide a good balance of speed and memory. If 12 GB VRAM is available, Q6_K (1.98 GB) offers higher precision. Larger Q8_0, BF16, and F16 require 2.5 GB or more and are best for systems with 16 GB+ VRAM.
What GPU do you need for Granite Speech 4.1 2B NAR?
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 Q8_0 (2.5 GB file, about 3.3 GB in use).
- 6 GB VRAM: the best quant that fits is F16 (4.5 GB file, about 5.6 GB in use).
Which Granite Speech 4.1 2B NAR quant should you choose?
Q4_K_M is the usual starting point for Granite Speech 4.1 2B NAR: it balances file size and quality well. Here it is a 1.6 GB file.
The Q8_0 file is 2.5 GB, about 1.6 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.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about Granite Speech 4.1 2B NAR
How much VRAM does Granite Speech 4.1 2B NAR need?
About 2.2 GB at Q4_K_M with an 8K context window. That is the 1.6 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 Granite Speech 4.1 2B NAR on an 8 GB GPU?
Yes, with F16: the 4.5 GB file needs about 5.6 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of Granite Speech 4.1 2B NAR GGUF?
It depends on the quant: Q4_K_M is 1.6 GB, Q5_K_M is 1.8 GB, Q6_K is 2.0 GB, Q8_0 is 2.5 GB, BF16 is 4.5 GB, F16 is 4.5 GB.
Where can I download Granite Speech 4.1 2B NAR GGUF?
The files are on Hugging Face in the handy-computer/granite-speech-4.1-2b-nar-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 Granite Speech 4.1 2B NAR 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 Granite Speech 4.1 2B NAR on 8 GB VRAM?
Yes, using Q4_K_M or Q5_K_M quantizations keeps VRAM usage below 8 GB, with estimated memory around 2.2–2.5 GB at 8k context.
Does the model support speaker diarization?
No, the NAR version does not provide speaker labels or word‑level timing.
What license applies to this model?
The model is released under the Apache 2.0 license.