/open-llms

📋 A list of open LLMs available for commercial use.

Apache License 2.0Apache-2.0

Open LLMs

These LLMs (Large Language Models) are all licensed for commercial use (e.g., Apache 2.0, MIT, OpenRAIL-M). Contributions welcome!

Language Model Release Date Checkpoints Paper/Blog Params (B) Context Length Licence Try it
T5 2019/10 T5 & Flan-T5, Flan-T5-xxl (HF) Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer 0.06 - 11 512 Apache 2.0 T5-Large
RWKV 4 2021/08 RWKV, ChatRWKV The RWKV Language Model (and my LM tricks) 0.1 - 14 infinity (RNN) Apache 2.0
GPT-NeoX-20B 2022/04 GPT-NEOX-20B GPT-NeoX-20B: An Open-Source Autoregressive Language Model 20 2048 Apache 2.0
YaLM-100B 2022/06 yalm-100b Yandex publishes YaLM 100B, the largest GPT-like neural network in open source 100 1024 Apache 2.0
UL2 2022/10 UL2 & Flan-UL2, Flan-UL2 (HF) UL2 20B: An Open Source Unified Language Learner 20 512, 2048 Apache 2.0
Bloom 2022/11 Bloom BLOOM: A 176B-Parameter Open-Access Multilingual Language Model 176 2048 OpenRAIL-M v1
ChatGLM 2023/03 chatglm-6b ChatGLM, Github 6 2048 Custom Free with some usage restriction (might require registration)
Cerebras-GPT 2023/03 Cerebras-GPT Cerebras-GPT: A Family of Open, Compute-efficient, Large Language Models (Paper) 0.111 - 13 2048 Apache 2.0 Cerebras-GPT-1.3B
Open Assistant (Pythia family) 2023/03 OA-Pythia-12B-SFT-8, OA-Pythia-12B-SFT-4, OA-Pythia-12B-SFT-1 Democratizing Large Language Model Alignment 12 2048 Apache 2.0 Pythia-2.8B
Pythia 2023/04 pythia 70M - 12B Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling 0.07 - 12 2048 Apache 2.0
Dolly 2023/04 dolly-v2-12b Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM 3, 7, 12 2048 MIT
StableLM-Alpha 2023/04 StableLM-Alpha Stability AI Launches the First of its StableLM Suite of Language Models 3 - 65 4096 CC BY-SA-4.0
FastChat-T5 2023/04 fastchat-t5-3b-v1.0 We are excited to release FastChat-T5: our compact and commercial-friendly chatbot! 3 512 Apache 2.0
DLite 2023/05 dlite-v2-1_5b Announcing DLite V2: Lightweight, Open LLMs That Can Run Anywhere 0.124 - 1.5 1024 Apache 2.0 DLite-v2-1.5B
h2oGPT 2023/05 h2oGPT Building the World’s Best Open-Source Large Language Model: H2O.ai’s Journey 12 - 20 256 - 2048 Apache 2.0
MPT-7B 2023/05 MPT-7B, MPT-7B-Instruct Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs 7 84k (ALiBi) Apache 2.0, CC BY-SA-3.0
RedPajama-INCITE 2023/05 RedPajama-INCITE Releasing 3B and 7B RedPajama-INCITE family of models including base, instruction-tuned & chat models 3 - 7 2048 Apache 2.0 RedPajama-INCITE-Instruct-3B-v1
OpenLLaMA 2023/05 open_llama_3b, open_llama_7b, open_llama_13b OpenLLaMA: An Open Reproduction of LLaMA 3, 7 2048 Apache 2.0 OpenLLaMA-7B-Preview_200bt
Falcon 2023/05 Falcon-180B, Falcon-40B, Falcon-7B The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only 180, 40, 7 2048 Apache 2.0
GPT-J-6B 2023/06 GPT-J-6B, GPT4All-J GPT-J-6B: 6B JAX-Based Transformer 6 2048 Apache 2.0
MPT-30B 2023/06 MPT-30B, MPT-30B-instruct MPT-30B: Raising the bar for open-source foundation models 30 8192 Apache 2.0, CC BY-SA-3.0 MPT 30B inference code using CPU
LLaMA 2 2023/06 LLaMA 2 Weights  Llama 2: Open Foundation and Fine-Tuned Chat Models 7 - 70 4096 Custom Free if you have under 700M users and you cannot use LLaMA outputs to train other LLMs besides LLaMA and its derivatives HuggingChat
ChatGLM2 2023/06 chatglm2-6b ChatGLM2-6B, Github 6 32k Custom Free with some usage restriction (might require registration)
XGen-7B 2023/06 xgen-7b-4k-base, xgen-7b-8k-base Long Sequence Modeling with XGen 7 4096, 8192 Apache 2.0
Jais-13b 2023/08 jais-13b, jais-13b-chat Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models 13 2048 Apache 2.0
OpenHermes 2023/09 OpenHermes-7B, OpenHermes-13B Nous Research 7, 13 4096 MIT OpenHermes-V2 Finetuned on Mistral 7B
OpenLM 2023/09 OpenLM 1B, OpenLM 7B  Open LM: a minimal but performative language modeling (LM) repository 1, 7 2048 MIT
Mistral 7B 2023/09 Mistral-7B-v0.1, Mistral-7B-Instruct-v0.1 Mistral 7B 7 4096-16K with Sliding Windows Apache 2.0 Mistral Transformer
ChatGLM3 2023/10 chatglm3-6b, chatglm3-6b-base, chatglm3-6b-32k, chatglm3-6b-128k ChatGLM3 6 8192, 32k, 128k Custom Free with some usage restriction (might require registration)
Skywork 2023/10 Skywork-13B-Base, Skywork-13B-Math Skywork 13 4096 Custom Free with usage restriction and models trained on Skywork outputs become Skywork derivatives, subject to this license.
Jais-30b 2023/11 jais-30b-v1, jais-30b-chat-v1 Jais-30B: Expanding the Horizon in Open-Source Arabic NLP 30 2048 Apache 2.0
Zephyr 2023/11 Zephyr 7B Website 7 8192 Apache 2.0
DeepSeek 2023/11 deepseek-llm-7b-base, deepseek-llm-7b-chat, deepseek-llm-67b-base, deepseek-llm-67b-chat Introducing DeepSeek LLM, 7, 67 4096 Custom Free with usage restriction and models trained on DeepSeek outputs become DeepSeek derivatives, subject to this license.
Mistral 7B v0.2 2023/12 Mistral-7B-v0.2, Mistral-7B-Instruct-v0.2 La Plateforme 7 32k Apache 2.0
Mixtral 8x7B v0.1 2023/12 Mixtral-8x7B-v0.1, Mixtral-8x7B-Instruct-v0.1 Mixtral of experts 46.7 32k Apache 2.0
LLM360 Amber 2023/12 Amber, AmberChat, AmberSafe Introducing LLM360: Fully Transparent Open-Source LLMs 6.7 2048 Apache 2.0
SOLAR 2023/12 Solar-10.7B Upstage 10.7 4096 apache-2.0
phi-2 2023/12 phi-2 2.7B Microsoft 2.7 2048 MIT
FLOR 2023/12 FLOR-760M, FLOR-1.3B, FLOR-1.3B-Instructed, FLOR-6.3B, FLOR-6.3B-Instructed FLOR-6.3B: a chinchilla-compliant model for Catalan, Spanish and English 0.76, 1.3, 6.3 2048 Apache 2.0 with usage restriction inherited from BLOOM
RWKV 5 v2 2024/01 rwkv-5-world-0.4b-2, rwkv-5-world-1.5b-2, rwkv-5-world-3b-2, rwkv-5-world-3b-2(16k), rwkv-5-world-7b-2 RWKV 5 0.4, 1.5, 3, 7 unlimited(RNN), trained on 4096 (and 16k for 3b) Apache 2.0
OLMo 2024/02 OLMo 1B, OLMo 7B, OLMo 7B Twin 2T AI2 1,7 2048 Apache 2.0
Qwen1.5 2024/02 Qwen1.5-7B, Qwen1.5-7B-Chat, Qwen1.5-14B, Qwen1.5-14B-Chat, Qwen1.5-72B, Qwen1.5-72B-Chat Introducing Qwen1.5 7, 14, 72 32k Custom Free if you have under 100M users and you cannot use Qwen outputs to train other LLMs besides Qwen and its derivatives
LWM 2024/02 LWM-Text-Chat-128K, LWM-Text-Chat-256K, LWM-Text-Chat-512K, LWM-Text-Chat-1M, LWM-Text-128K, LWM-Text-256K, LWM-Text-512K, LWM-Text-1M Large World Model (LWM) 7 128k, 256k, 512k, 1M LLaMA 2 license
Jais-30b v3 2024/03 jais-30b-v3, jais-30b-chat-v3 Jais 30b v3 30 8192 Apache 2.0
Gemma 2024/02 Gemma 7B, Gemma 7B it, Gemma 2B, Gemma 2B it Technical report 2-7 8192 Gemma Terms of Use Free with usage restriction and models trained on Gemma outputs become Gemma derivatives, subject to this license.
Grok-1 2024/03 Grok-1 Open Release of Grok-1 314 8192 Apache 2.0
Qwen1.5 MoE 2024/03 Qwen1.5-MoE-A2.7B, Qwen1.5-MoE-A2.7B-Chat Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters 14.3 8192 Custom Free if you have under 100M users and you cannot use Qwen outputs to train other LLMs besides Qwen and its derivatives
Jamba 0.1 2024/03 Jamba-v0.1 Introducing Jamba: AI21's Groundbreaking SSM-Transformer Model 52 256k Apache 2.0
Qwen1.5 32B 2024/04 Qwen1.5-32B, Qwen1.5-32B-Chat Qwen1.5-32B: Fitting the Capstone of the Qwen1.5 Language Model Series 32 32k Custom Free if you have under 100M users and you cannot use Qwen outputs to train other LLMs besides Qwen and its derivatives
Mamba-7B 2024/04 mamba-7b-rw Toyota Research Institute 7 unlimited(RNN), trained on 2048 Apache 2.0
Mixtral8x22B v0.1 2024/04 Mixtral-8x22B-v0.1, Mixtral-8x22B-Instruct-v0.1 Cheaper, Better, Faster, Stronger 141 64k Apache 2.0
Llama 3 2024/04 Llama-3-8B, Llama-3-8B-Instruct, Llama-3-70B, Llama-3-70B-Instruct, Llama-Guard-2-8B Introducing Meta Llama 3, Meta Llama 3 8, 70 8192 Meta Llama 3 Community License Agreement Free if you have under 700M users and you cannot use LLaMA 3 outputs to train other LLMs besides LLaMA 3 and its derivatives
Phi-3 Mini 2024/04 Phi-3-mini-4k-instruct, Phi-3-mini-128k-instruct Introducing Phi-3, Technical Report 3.8 4096, 128k MIT
OpenELM 2024/04 OpenELM-270M, OpenELM-270M-Instruct, OpenELM-450M, OpenELM-450M-Instruct, OpenELM-1_1B, OpenELM-1_1B-Instruct, OpenELM-3B, OpenELM-3B-Instruct OpenELM: An Efficient Language Model Family with Open Training and Inference Framework 0.27, 0.45, 1.1, 3 2048 Custom open license No usage or training restrictions
Snowflake Arctic 2024/04 snowflake-arctic-base, snowflake-arctic-instruct Snowflake Arctic: The Best LLM for Enterprise AI — Efficiently Intelligent, Truly Open 480 4096 Apache 2.0
Qwen1.5 110B 2024/04 Qwen1.5-110B, Qwen1.5-110B-Chat Qwen1.5-110B: The First 100B+ Model of the Qwen1.5 Series 110 32k Custom Free if you have under 100M users and you cannot use Qwen outputs to train other LLMs besides Qwen and its derivatives
RWKV 6 v2.1 2024/05 rwkv-6-world-1.6b-2.1, rwkv-6-world-3b-2.1, rwkv-6-world-7b-2.1 RWKV 6 1.6, 3, 7 unlimited(RNN), trained on 4096 Apache 2.0
DeepSeek-V2 2024/05 DeepSeek-V2, DeepSeek-V2-Chat DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model 236 128k Custom Free with usage restriction and models trained on DeepSeek outputs become DeepSeek derivatives, subject to this license.
Fugaku-LLM 2024/05 Fugaku-LLM-13B, Fugaku-LLM-13B-instruct Release of "Fugaku-LLM" – a large language model trained on the supercomputer "Fugaku" 13 2048 Custom Free with usage restrictions
Falcon 2 2024/05 falcon2-11B Meet Falcon 2: TII Releases New AI Model Series, Outperforming Meta’s New Llama 3 11 8192 Custom Apache 2.0 with mild acceptable use policy
Yi-1.5 2024/05 Yi-1.5-6B, Yi-1.5-6B-Chat, Yi-1.5-9B, Yi-1.5-9B-Chat, Yi-1.5-34B, Yi-1.5-34B-Chat Yi-1.5 6, 9, 34 4096 Apache 2.0
DeepSeek-V2-Lite 2024/05 DeepSeek-V2-Lite, DeepSeek-V2-Lite-Chat DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model 16 32k Custom Free with usage restriction and models trained on DeepSeek outputs become DeepSeek derivatives, subject to this license.
Phi-3 small/medium 2024/05 Phi-3-mini-4k-instruct, Phi-3-mini-128k-instruct, Phi-3-medium-4k-instruct, Phi-3-medium-128k-instruct New models added to the Phi-3 family, available on Microsoft Azure, Technical Report 7, 14 4096, 128k MIT

Open LLMs for code

Language Model Release Date Checkpoints Paper/Blog Params (B) Context Length Licence Try it
SantaCoder 2023/01 santacoder SantaCoder: don't reach for the stars! 1.1 2048 OpenRAIL-M v1 SantaCoder
CodeGen2 2023/04 codegen2 1B-16B CodeGen2: Lessons for Training LLMs on Programming and Natural Languages 1 - 16 2048 Apache 2.0
StarCoder 2023/05 starcoder StarCoder: A State-of-the-Art LLM for Code, StarCoder: May the source be with you! 1.1-15 8192 OpenRAIL-M v1
StarChat Alpha 2023/05 starchat-alpha Creating a Coding Assistant with StarCoder 16 8192 OpenRAIL-M v1
Replit Code 2023/05 replit-code-v1-3b Training a SOTA Code LLM in 1 week and Quantifying the Vibes — with Reza Shabani of Replit 2.7 infinity? (ALiBi) CC BY-SA-4.0 Replit-Code-v1-3B
CodeT5+ 2023/05 CodeT5+ CodeT5+: Open Code Large Language Models for Code Understanding and Generation 0.22 - 16 512 BSD-3-Clause Codet5+-6B
XGen-7B 2023/06 XGen-7B-8K-Base Long Sequence Modeling with XGen: A 7B LLM Trained on 8K Input Sequence Length 7 8192 Apache 2.0
CodeGen2.5 2023/07 CodeGen2.5-7B-multi CodeGen2.5: Small, but mighty 7 2048 Apache 2.0
DeciCoder-1B 2023/08 DeciCoder-1B Introducing DeciCoder: The New Gold Standard in Efficient and Accurate Code Generation 1.1 2048 Apache 2.0 DeciCoder Demo
Code Llama 2023/08 Inference Code for CodeLlama models  Code Llama: Open Foundation Models for Code 7 - 34 4096 Custom Free if you have under 700M users and you cannot use LLaMA outputs to train other LLMs besides LLaMA and its derivatives HuggingChat

Open LLM datasets for pre-training

Name Release Date Paper/Blog Dataset Tokens (T) License
RedPajama 2023/04 RedPajama, a project to create leading open-source models, starts by reproducing LLaMA training dataset of over 1.2 trillion tokens RedPajama-Data 1.2 Apache 2.0
starcoderdata 2023/05 StarCoder: A State-of-the-Art LLM for Code starcoderdata 0.25 Apache 2.0

Open LLM datasets for instruction-tuning

Name Release Date Paper/Blog Dataset Samples (K) License
OIG (Open Instruction Generalist) 2023/03 THE OIG DATASET OIG 44,000 Apache 2.0
databricks-dolly-15k 2023/04 Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM databricks-dolly-15k 15 CC BY-SA-3.0
MPT-7B-Instruct 2023/05 Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs dolly_hhrlhf 59 CC BY-SA-3.0

Open LLM datasets for alignment-tuning

Name Release Date Paper/Blog Dataset Samples (K) License
OpenAssistant Conversations Dataset 2023/04 OpenAssistant Conversations - Democratizing Large Language Model Alignment oasst1 161 Apache 2.0

Evals on open LLMs

What do the licences mean?

  • Apache 2.0: Allows users to use the software for any purpose, to distribute it, to modify it, and to distribute modified versions of the software under the terms of the license, without concern for royalties.
  • MIT: Similar to Apache 2.0 but shorter and simpler. Also, in contrast to Apache 2.0, does not require stating any significant changes to the original code.
  • CC BY-SA-4.0: Allows (i) copying and redistributing the material and (ii) remixing, transforming, and building upon the material for any purpose, even commercially. But if you do the latter, you must distribute your contributions under the same license as the original. (Thus, may not be viable for internal teams.)
  • OpenRAIL-M v1: Allows royalty-free access and flexible downstream use and sharing of the model and modifications of it, and comes with a set of use restrictions (see Attachment A)
  • BSD-3-Clause: This version allows unlimited redistribution for any purpose as long as its copyright notices and the license's disclaimers of warranty are maintained.

Disclaimer: The information provided in this repo does not, and is not intended to, constitute legal advice. Maintainers of this repo are not responsible for the actions of third parties who use the models. Please consult an attorney before using models for commercial purposes.

Improvements

  • Complete entries for context length, and check entries with ?
  • Add number of tokens trained? (see considerations)
  • Add (links to) training code?
  • Add (links to) eval benchmarks?