LFM2.5 2.6B DSpark Draft v1 needs about 0.8 GB of memory at Q4_K_M (a 0.2 GB file) with an 8K context window, so it fits on a 4 GB GPU. This repository has 3 quants, from 0.2 GB (Q4_K_M) to 0.7 GB (F16). Longer context needs more: at 32K tokens the KV cache alone adds about 0.3 GB.
LFM2.5 2.6B DSpark Draft v1 – a lightweight draft sidecar for faster inference
LFM2.5 2.6B DSpark Draft v1 is a draft-only model built on the LFM2.5 architecture. It contains 327.7 M draft parameters and is designed for speculative decoding with the target LFM2.5‑2.6B model. The draft is 5 attention layers deep and shares embeddings with the target.
Good at
- Speculative decoding draft
- Fast inference with LFM2.5‑2.6B target
- Low memory footprint
Limits
- Requires paired LFM2.5‑2.6B‑GGUF target
- No vision or multimodal support
- License is other – check usage rights
Best for
- Speed‑critical text generation
- Edge deployment with limited GPU memory
- Batch‑size‑1 inference
Who it’s for. Developers building on LFM2.5 models who need faster generation without sacrificing quality. Useful for applications that can tolerate a draft sidecar and have access to the target model.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 0.2 GB | 0.8 GB | Fits | Get file → | |
| Q8_0 | 0.4 GB | 0.9 GB | Fits | Get file → | |
| F16 | 0.7 GB | 1.2 GB | Fits | Get file → |
For 8 GB VRAM, Q4_K_M (0.2 GB) is recommended; it uses ~0.8 GB at 8 k context. If 12 GB is available, Q8_0 (0.36 GB) gives higher precision with ~0.9 GB usage. F16 (0.66 GB) is only suitable for larger memory budgets.
What GPU do you need for LFM2.5 2.6B DSpark Draft v1?
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 F16 (0.7 GB file, about 1.2 GB in use).
Which LFM2.5 2.6B DSpark Draft v1 quant should you choose?
Q4_K_M is the usual starting point for LFM2.5 2.6B DSpark Draft v1: it balances file size and quality well. Here it is a 0.2 GB file.
The Q8_0 file is 0.4 GB, about 1.8 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 LFM2.5 2.6B DSpark Draft v1
How much VRAM does LFM2.5 2.6B DSpark Draft v1 need?
About 0.8 GB at Q4_K_M 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 LFM2.5 2.6B DSpark Draft v1 on an 8 GB GPU?
Yes, with F16: the 0.7 GB file needs about 1.2 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of LFM2.5 2.6B DSpark Draft v1 GGUF?
It depends on the quant: Q4_K_M is 0.2 GB, Q8_0 is 0.4 GB, F16 is 0.7 GB.
What context length does LFM2.5 2.6B DSpark Draft v1 support?
The GGUF header lists a trained context length of 128,000 tokens (125K). The KV cache grows with the context you use: about 0.1 GB at 8K and 0.3 GB at 32K.
Where can I download LFM2.5 2.6B DSpark Draft v1 GGUF?
The files are on Hugging Face in the LiquidAI/LFM2.5-2.6B-DSpark-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 LFM2.5 2.6B DSpark Draft v1 use?
Hugging Face lists the license as other. 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 this draft model on 8 GB VRAM?
Yes, the Q4_K_M quant uses about 0.8 GB at 8 k context, fitting within 8 GB VRAM.
Does this model support vision or multimodal inputs?
No, it is a text‑only draft sidecar and has no vision capabilities.
Do I need a special build of llama.cpp to use it?
Yes, llama.cpp must be built with DSpark support for LFM2 targets to enable speculative decoding.