FA

Falcon

FalconGGUF by mradermacherUpdated 2024-05-06
Download on Hugging Face
Parameters41.8B
Context2K
Layers60
Architecturefalcon
Smallest quant9.7 GB
Quick answer

Falcon needs about 27.5 GB of memory at Q4_K_M (a 26.0 GB file) with an 8K context window, so it fits on a 32 GB GPU. This repository has 21 quants, from 9.7 GB (IQ1_S) to 35.0 GB (Q6_K).

Falcon-40B-Instruct, a 40B parameter causal decoder-only model fine‑tuned for instruction following

Falcon-40B-Instruct is a 40B parameter model built by TII and fine‑tuned on a mix of Baize and RefinedWeb. It is released under the Apache 2.0 license and supports English and French. The model is optimized for inference with FlashAttention and multi‑query attention.

Good at

  • Instruction‑driven chat
  • Large context handling
  • High‑quality text generation

Limits

  • Requires 85‑100 GB VRAM for full‑precision inference
  • No vision or multimodal support
  • License allows commercial use but requires compliance with Apache 2.0

Best for

  • Chatbot back‑ends
  • Content generation pipelines
  • Research on large‑scale instruction models

Who it’s for. Developers and researchers needing a large, instruction‑tuned decoder‑only model. The model is suitable for production chat services when adequate risk mitigation is in place.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ1_S9.7 GB 10.5 GB
Too big Get file →
IQ1_M10.0 GB 10.8 GB
Too big Get file →
IQ2_XXS11.9 GB 12.6 GB
Too big Get file →
IQ2_XS13.1 GB 13.9 GB
Too big Get file →
IQ2_S13.9 GB 14.7 GB
Too big Get file →
IQ2_M15.0 GB 15.8 GB
Too big Get file →
Q2_K16.3 GB 17.0 GB
Too big Get file →
IQ3_XXS16.9 GB 17.7 GB
Too big Get file →
IQ3_XS18.3 GB 19.0 GB
Too big Get file →
IQ3_S18.8 GB 19.6 GB
Too big Get file →
Q3_K_S18.8 GB 19.6 GB
Too big Get file →
IQ3_M19.7 GB 20.4 GB
Too big Get file →
Q3_K_M20.6 GB 21.3 GB
Too big Get file →
Q3_K_L22.1 GB 22.8 GB
Too big Get file →
IQ4_XS22.5 GB 23.3 GB
Too big Get file →
Q4_023.9 GB 24.7 GB
Too big Get file →
Q4_K_S24.3 GB 25.1 GB
Too big Get file →
Q4_K_M26.0 GB 26.7 GB
Too big Get file →
Q5_K_S29.5 GB 30.2 GB
Too big Get file →
Q5_K_M31.1 GB 31.9 GB
Too big Get file →
Q6_K35.0 GB 35.7 GB
Too big Get file →

For 8 GB VRAM, use IQ1_S (9.71 GB file, ~11.2 GB VRAM). For 16 GB VRAM, IQ2_XS (13.11 GB file, ~14.6 GB VRAM) or Q2_K (16.26 GB file, ~17.8 GB VRAM). For 24 GB VRAM, Q4_K_M (25.99 GB file, ~27.5 GB VRAM) provides a good balance of size and quality. Estimates are approximate.

What GPU do you need for Falcon?

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: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: the best quant that fits is IQ1_M (10.0 GB file, about 11.5 GB in use).
  • 16 GB VRAM: the best quant that fits is IQ2_S (13.9 GB file, about 15.4 GB in use).
  • 24 GB VRAM: the best quant that fits is IQ4_XS (22.5 GB file, about 24.0 GB in use).
  • 32 GB VRAM: the best quant that fits is Q5_K_S (29.5 GB file, about 31.0 GB in use).
  • 48 GB VRAM: the best quant that fits is Q6_K (35.0 GB file, about 36.5 GB in use).

Which Falcon quant should you choose?

Q4_K_M is the usual starting point for Falcon: it balances file size and quality well. Here it is a 26.0 GB file.

If memory is tight, the smallest file is IQ1_S at 9.7 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturefalcon
Parameters41.8B
Layers60
Trained context2,048 tokens
KV cache (fp16)120 KB per token

Source

Quantized bymradermacher
FamilyFalcon
Licenseapache-2.0
Downloads (HF)4,292
Last updated2024-05-06

Frequently asked questions about Falcon

How much VRAM does Falcon need?

About 27.5 GB at Q4_K_M with an 8K context window. That is the 26.0 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 Falcon on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 11.2 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Falcon GGUF?

It depends on the quant: IQ1_S is 9.7 GB, IQ1_M is 10.0 GB, IQ2_XXS is 11.9 GB, IQ2_XS is 13.1 GB, IQ2_S is 13.9 GB, IQ2_M is 15.0 GB and 15 more.

What context length does Falcon support?

The GGUF header lists a trained context length of 2,048 tokens (2K). The KV cache grows with the context you use: about 1.0 GB at 8K and 0.3 GB at 2K.

Where can I download Falcon GGUF?

The files are on Hugging Face in the mradermacher/falcon-40b-instruct-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 Falcon 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 Falcon-40B-Instruct on 8 GB VRAM?

Use the IQ1_S quant, which is 9.71 GB on disk and requires about 11.2 GB VRAM at 8k context.

Does Falcon-40B-Instruct support images?

No, it is a text‑only causal decoder‑only model with no vision capability.

What license applies to Falcon-40B-Instruct?

The model is released under the Apache 2.0 license, allowing commercial use with attribution.

Other GGUFs of falcon-40b-instruct

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More Falcon models

All Falcon →

Other large models

All large models →