Proton, the privacy-friendly internet service provider based in Europe, has jumped on the LLM bandwagon and released an LLM service last month called Lumo. Besides touting its privacy features, Proton's PR focuses on open source:
Unlike other AI assistants, my code is fully open source, so anyone can verify that it’s private and secure — and that we never use your data to train the model. (source)
Elsewhere on the Proton website, we find a claim that Lumo is "based upon open-source language models". A comparison table shows Lumo alongside some other LLM assistants, with a feature "Opens source code for the public" prominently checked for Lumo and DeepSeek. Factcheck: for DeepSeek that is definitely not the case; at best it is an open weights model and very little is known about its source code. Does Lumo fare any better?
It turns out that Lumo hits a new record in openness: it is the least open "open" model that we have ever added to our index. One of our reasons for inclusion is an openness claim: a model provider that calls a model "open" or "open source" or a variation on that. Lumo is "open" in that sense (Proton calls it open source) but in no other way. Nothing about it is currently open.
Here is a comparison of the openness status of Lumo and two well-known other models in the space: DeepSeek (included in Proton's own comparison), and OLMo from AllenAI, the current leading openness champion.
Parameter descriptions:
Base Model Data
Are datasources for training the base model comprehensively documented and made available? In case a distinction between base (foundation) and end (user) model is not applicable, this mirrors the end model data entries.
End User Model Data
Are datasources for training the model that the end user interacts with comprehensively documented and made available?
Base Model Weights
Are the weights of the base models made freely available? In case a distinction between base (foundation) and end (user) model is not applicable, this mirrors the end model data entries.
End User Model Weights
Are the weights of the model that the end user interacts with made freely available?
Training Code
Is the source code of dataset processing, model training and tuning comprehensively made available?
Code Documentation
Is the source code of datasource processing, model training and tuning comprehensively documented?
Hardware Architecture
Is the hardware architecture used for datasource processing and model training comprehensively documented?
Preprint
Are archived preprint(s) are available that detail all major parts of the system including datasource processing, model training and tuning steps?
Paper
Are peer-reviewed scientific publications available that detail all major parts of the system including datasource processing, model training and tuning steps?
Modelcard
Is a model card available in standardized format that provides comprehensive insight on model architecture, training, fine-tuning, and evaluation?
Datasheet
Is a datasheet as defined in "Datasheets for Datasets" (Gebru et al. 2021) available?
Package
Is a packaged release of the model available on a software repository (e.g. a Python Package Index, Homebrew)?
API and Meta Prompts
Is an API available that provides unrestricted access to the model (other than security and CDN restrictions)? If applicable, this entry also collects information on the use and availability of meta prompts.
Licenses
Is the project fully covered by Open Source Initiative (OSI)-approved licenses, including all data sources and training pipeline code?
YuLan by Gaoling School of Artificial Intelligence
YuLan-Mini
BLOOMZ by BigScience Workshop
BLOOM
Poro by Silo AI and TurkuNLP and High Performance Language Technologies (HPLT)
Poro-34B
Open Assistant by LAION
Pythia-12B
mT0 by BigScience Workshop
mT5-XXL
Whisper by OpenAI
Whisper-large-v3
Pythia by EleutherAI and Together Computer
Pythia-6.9B
Amber by LLM360
Amber
K2 by LLM360
K2
SmolLM by HuggingFace
SmolLM2-1.7B
OpenChat by OpenChat
Meta-Llama-3-8B
Arabic StableLM by StabilityAI
StableLM-2-1.6B
Teuken by OpenGPT-X
Teuken-7B-base
Instella by AMD
Instella-3B
Dolly by Databricks
Pythia-12B
Tülu by Ai2
Llama-3.1-405B
T5 by Google AI
T5
RedPajama by Together Computer
RedPajama-INCITE-7B-Base
Neo by Multimodal Art Projection
Neo-7B
BERT by Google AI
BERT
AquilaChat by Beijing Academy of Artificial Intelligence
Aquila2-70B-Expr
Eurus by OpenBMB
Mixtral-8x22B-v0.1
DeepSeek V3 by DeepSeek
DeepSeek-V3-Base
Yi by 01.AI
Yi-34B
Salamandra by Barcelona Supercomputing Center
Salamandra-7B
NeuralChat by Intel
Mistral-7B-v0.1
MPT by Databricks
MPT-30B
Lucie by OpenLLM-France
Lucie-7B
Guru by LLM360
Guru-32B
GPT-SW3 by AI Sweden
GPT-SW3-6.7B-V2
GPT-NeoXT by Together Computer
GPT-NeoX-20B
Fietje by Bram Vanroy
Phi-2
BTLM by Cerebras
BTLM-3B-8K-Base
Pharia by Aleph Alpha Research
Pharia-1-LLM-7B
minChatGPT by Ethan Yanjia Li
GPT2-Medium
Xwin-LM by Xwin-LM
Llama-2-13B
Vicuna by LMSYS
Vicuna-13B
Phi by Microsoft
Phi-4
OpenELM by Apple
OpenELM-3B
Occiglot by Occiglot
Occiglot-7B-EU5
Mistral by Mistral AI
Mistral-Large-2411
GLM by Zhipu AI
GLM-4-32B-0414
Falcon by Technology Innovation Institute
Falcon3-10B-Base
Minerva by Sapienza Natural Language Processing Group
Minerva-7B-base-v1.0
DeepSeek R1 by DeepSeek
DeepSeek-V3-Base
Zephyr by HuggingFace
Mixtral-8x22B-v0.1
WizardLM by Microsoft and Peking University
LLaMA-7B
SynLogic by Minimax AI
SynLogic-32B
InternLM by Shanghai AI Laboratory
InternLM3-8B
GPT OSS by OpenAI
GPT-OSS-120B
CT-LLM by Multimodal Art Projection
CT-LLM-Base
Mistral NeMo by Mistral AI and NVIDIA
Mistral-NeMo-12B-Base
Saul by Equall
Mixtral-8x22B-v0.1
Qwen by Alibaba
Qwen3-235B-A22B-Base
Kimi K2 by Moonshot AI
Kimi K2 Base
Granite by IBM
Granite-3.3-8B-Base
MiMo by Xiaomi
MiMo-7B-Base
Airoboros by Jon Durbin
Qwen1.5-110B
XBai-04 by Yuan Shi Technology
Qwen3-32B
Starling by NexusFlow
Llama-2-13B
Gemma by Google AI
Gemma-3-27B-PT
Geitje by Bram Vanroy
Mistral-7B-v0.1
BELLE by KE Technologies
Llama-2-13B
Llama 4 by Meta
Llama-4-Maverick-17B-128E
dots.llm1 by RedNote
dots.llm1.base
Marco by Alibaba
Marco-LLM-GLO
Viking by Silo AI and TurkuNLP and High Performance Language Technologies (HPLT)
Viking-33B
Llama 3.1 by Meta
Llama-3.1-405B
OpenMoE by Zheng Zian
OpenMoE-8B
LongAlign by Zhipu AI
Llama-2-13B
UltraLM by OpenBMB
Llama-13B
Command-R by Cohere AI
C4AI-Command-R-V01
Stanford Alpaca by Stanford University CRFM
Llama-7B
StripedHyena by Together Computer
StripedHyena-Hessian-7B
Claire by OpenLLM-France
Falcon-7B
Llama 3.3 by Meta
Llama-3.3-70B
Stable Beluga by StabilityAI
Llama-2-70B
Solar by Upstage AI
Mistral-7B-v0.1
RWKV by BlinkDL/RWKV
RWKV-x070-Pile-1.47B-ctx4096
Persimmon by Adept AI Labs
Persimmon-8B-Base
OPT by Meta
OPT-30B
Nanbeige by Nanbeige LLM lab
Nanbeige2-16B
Jais by G42
Llama-2-70B
Infinity-Instruct by Beijing Academy of Artificial Intelligence
Llama-3.1-70B
H2O-Danube by H2O.ai
H2O-Danube3.1-4B-Chat
FastChat-T5 by LMSYS
Flan-T5-XL
Crystal by LLM360
Crystal
BitNet by Microsoft
BitNet-b1.58-2B4T
Baichuan by Baichuan Intelligent Technology
Baichuan2-13B-Base
StableVicuna by CarperAI
LLaMA-13B
Llama 3 by Meta
Meta-Llama-3-70B
Llama 2 by Meta
Llama-2-70B
Koala by BAIR
Llama-13B
XGen by Salesforce
XGen-Small-9B-Base-R
Hunyuan by Tencent
Hunyuan-A52B-Pretrain
Gemma Japanese by Google AI
Gemma-2-2B
Snowflake Arctic by Snowflake
Snowflake-Arctic-Base
Llama-Sherkala by G42
Llama-3.1-8B
DeepHermes by Nous Research
Llama-3.1-8B
Minimax-Text by Minimax AI
MiniMax-Text-01
Lumo AI by Proton
Undisclosed
In the past, we've gotten model providers like Mistral retreat from using the term "open source" and instead use the more precise "open weights". But Proton will need to do more work to get there, as they don't even reach the middling openness levels of Mistral and other model providers evasive about training data.
On the sunny side: the only way from here is up. Literally disclosing anything, from training data to model weights to model architecture to data sheets, will be an improvement over the current status. We will be following its rise with great interest.
It may be that Lumo is a bespoke fine-tuned version of an open weights model like Mistral, OLMo or Llama. It may be a wrapper for an API from Anthropic or another provier. It may be all of these things together in a trenchcoat. The point is: we would know if it were truly open source.
