LLM EXPLORER 66,613 MODELS INDEXED

Falcon 40B Instruct 8bit by ichitaka

By ichitaka · 12 downloads
HF Repo ↗ Copy URL

Falcon 40B Instruct 8bit is an open-source language model by ichitaka. Features: 40b LLM, VRAM: 41.8GB, License: apache-2.0, Quantized, Instruction-Based, LLM Explorer Score: 0.07.

  Arxiv:1911.02150   Arxiv:2005.14165   Arxiv:2104.09864   Arxiv:2205.14135   8-bit   8bit   Custom code Dataset:tiiuae/falcon-refinedw...   En   Instruct   Pytorch   Quantized   Refinedweb   Region:us   Sharded
Model Card on HF 🤗: https://huggingface.co/ichitaka/falcon-40b-instruct-8bit 

Falcon 40B Instruct 8bit Parameters and Internals

Model Type 
Causal decoder-only
Use Cases 
Areas:
Chatbots, Instruction-following systems
Primary Use Cases:
Ready-to-use chat/instruct model based on Falcon-40B
Limitations:
Mostly trained on English data, may not generalize to other languages
Considerations:
Ensure appropriate guardrails are in place for production systems.
Additional Notes 
This model is a 40B parameter instruction model designed for use with GPUs using bitsandbytes for quantization.
Supported Languages 
English (primary), French (secondary)
Training Details 
Data Sources:
Baize, RefinedWeb
Data Volume:
150M tokens
Methodology:
Finetuned on a mixture of data
Context Length:
2048
Hardware Used:
64 A100 40GB GPUs
Model Architecture:
Causal decoder-only model with rotary positional embeddings and FlashAttention, multiquery attention, optimized MLP layer with single layer norm, parallel attention/MLP.
Safety Evaluation 
Risk Categories:
Bias - related to language; stereotypes from online data
Ethical Considerations:
Users should implement guardrails and assess risks for production use.
Responsible Ai Considerations 
Fairness:
Model primarily trained on English data, which may introduce bias and stereotypes.
Mitigation Strategies:
Developers should consider precautions for production deployment.
Input Output 
Accepted Modalities:
text
Performance Tips:
Recommended for use on systems with adequate GPU resources.
LLM NameFalcon 40B Instruct 8bit
Repository 🤗https://huggingface.co/ichitaka/falcon-40b-instruct-8bit 
Base Model(s)  Medfalcon 40B Lora   nmitchko/medfalcon-40b-lora
Model Size40b
Required VRAM41.8 GB
Updated2026-08-07
Maintainerichitaka
Model TypeRefinedWeb
Instruction-BasedYes
Model Files  10.0 GB: 1-of-5   9.8 GB: 2-of-5   9.9 GB: 3-of-5   9.8 GB: 4-of-5   2.3 GB: 5-of-5
Supported Languagesen
Quantization Type8bit
Model ArchitectureRWForCausalLM
Licenseapache-2.0
Model Max Length2048
Transformers Version4.30.0.dev0
Is Biased0
Tokenizer ClassPreTrainedTokenizerFast
Vocabulary Size65024
Torch Data Typefloat16

Best Alternatives to Falcon 40B Instruct 8bit

Best Alternatives
Context / RAM
Downloads
Likes
Falcon 40B Instruct GPTQ0K / 22.5 GB152197
...alcon 40B Instruct W4 G128 AWQ0K / 22.3 GB152
Falcon 40B Instruct GPTQ0K / 22.5 GB111
H2ogpt Oig Oasst1 Falcon 40B0K / 82.5 GB96
...truct GPTQ Inference Endpoints0K / 22.5 GB122
Note: green Score (e.g. "73.2") means that the model is better than ichitaka/falcon-40b-instruct-8bit.