What are the hardware requirements for Falcon 40B Instruct?

Falcon 40B Instruct requires approximately 83.6 GB of VRAM. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed Falcon 40B Instruct and how large is it?

Falcon 40B Instruct is developed by tiiuae, a model with 40b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

Where can I download or evaluate Falcon 40B Instruct?

Falcon 40B Instruct is hosted on Hugging Face and linked from this page. LLM Explorer also lists quantized variants and similar alternatives if available.

Falcon 40B Instruct by tiiuae — VRAM 83.6GB

Name: Falcon 40B Instruct
Author: tiiuae

Falcon 40B Instruct is an open-source language model by tiiuae. Features: 40b LLM, VRAM: 83.6GB, License: apache-2.0, Instruction-Based, LLM Explorer Score: 0.15.

Arxiv:1911.02150 Arxiv:2005.14165 Arxiv:2104.09864 Arxiv:2205.14135 Arxiv:2304.01196 Arxiv:2306.01116 Custom code Dataset:tiiuae/falcon-refinedw... Deploy:azure Deploy:sagemaker En Endpoints compatible Eval-results Falcon Instruct Pytorch Region:us Sharded

Model Card on HF 🤗: https://huggingface.co/tiiuae/falcon-40b-instruct

Falcon 40B Instruct Parameters and Internals

Model Type

Causal decoder-only

Use Cases

Areas:

chatbot, instruction-following

Applications:

chat, instruction-based interactions

Primary Use Cases:

ready-to-use chat/instruct model

Limitations:

Model mostly trained on English data, may not generalize well to other languages

Considerations:

Develop guardrails and take precautions for production use.

Additional Notes

Instruct model, not ideal for further finetuning. Optimized architecture for inference featuring FlashAttention and multiquery.

Supported Languages

English (primary), French (secondary)

Training Details

Data Sources:

Baize instruction dataset, RefinedWeb

Data Volume:

150M tokens from Baize mixed with 5% RefinedWeb

Methodology:

Finetuned on a mixture of chat data with 5% RefinedWeb

Context Length:

2048

Hardware Used:

64 A100 40GB GPUs on AWS SageMaker

Model Architecture:

Causal decoder-only with adaptations from GPT-3, including rotary embeddings, multiquery attention, FlashAttention, and a single layer norm with parallel attention/MLP

Input Output

Accepted Modalities:

text

LLM Name	Falcon 40B Instruct
Repository 🤗	https://huggingface.co/tiiuae/falcon-40b-instruct
Model Size	40b
Required VRAM	83.6 GB
Updated	2026-08-02
Maintainer	tiiuae
Model Type	falcon
Instruction-Based	Yes
Model Files	9.5 GB: 1-of-9 9.5 GB: 2-of-9 9.5 GB: 3-of-9 9.5 GB: 4-of-9 9.5 GB: 5-of-9 9.5 GB: 6-of-9 9.5 GB: 7-of-9 9.5 GB: 8-of-9 7.6 GB: 9-of-9
Supported Languages	en
Model Architecture	FalconForCausalLM
License	apache-2.0
Model Max Length	2048
Transformers Version	4.26.0
Is Biased	0
Tokenizer Class	PreTrainedTokenizerFast
Vocabulary Size	65024
Torch Data Type	bfloat16

Rank the Falcon 40B Instruct capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Best Alternatives to Falcon 40B Instruct

Best Alternatives	Context / RAM	Downloads	Likes
... Falcon 40B Instruct 4 Bit Bnb	2K / 23.9 GB	8	0
...Falcon 40B Instruct 4 Bit Gptq	2K / 22.3 GB	7	0