1 of 27

2024 in Open Models

Luca Soldaini, Nathan Lambert

2 of 27

2023

2024

Llama 1&2

Mistral

MPT

Falcon

Yi

Llama 3

Qwen 2.5

Deepseek

Cohere ⌘R

DBRX

OLMo

Mistral

LLM 360

DCLM

Gemma

3 of 27

4 of 27

Why Open Models?

the "Just use model APIs" argument

5 of 27

ResearchThrives Open Models

Modeling

Behavior

Inference

Evaluation

6 of 27

Open ModelsOutperform use cases

Retrieval (MTEB)

Edge AI

7 of 27

Community Stretches Beyond Models

https://qwenlm.github.io/blog/qwen2

8 of 27

CollaborationWhen we're free to participate, we can enhance each other's work in unanticipated ways.

https://opensource.com/open-source-way

9 of 27

2024: Open Source AI definition

https://opensource.org/ai

10 of 27

https://www.interconnects.ai/p/defining-open-source-ai

Good��Most of the definition is very standard. Weights need to be available, code for using the system must be released under an OSI-approved license, etc.

Compromise��Vague on data��... sufficiently detailed information about the data used to train the system, so that a skilled person can recreate a substantially equivalent system

11 of 27

We know way more than in 2023

We are more resource constrained than in 2023

12 of 27

2023

2024

Llama 1&2

Mistral

MPT

Falcon

Yi

Llama 3

Qwen 2.5

Deepseek

Cohere ⌘R

DBRX

OLMo

Mistral

LLM 360

DCLM

Gemma

Compute Rich Club�membership hotly debated

13 of 27

Pre-train

1k-∞

H100

Post-train

8-10k

H100

Inference

1-100

H100

Evaluation

0-100

H100

Research Budget

14 of 27

2023

2024

Llama 1&2

Mistral

MPT

Falcon

Yi

Llama 3

Qwen 2.5

Deepseek

Cohere ⌘R

DBRX

OLMo

Mistral

LLM 360

DCLM

Gemma

Fully Open Models�Hi! That's me.

15 of 27

Not just a checkpoint�the full recipe

https://arxiv.org/pdf/2409.02060

16 of 27

Any part of pipeline can be your starting point

17 of 27

Any part of pipeline can be your starting point

https://allenai.org/blog/olmo

18 of 27

Shameless plugAi2 fully open recipes

19 of 27

20 of 27

Open models in peril

It's not just compute

21 of 27

Harder to get data via crawling

Longpre et. al. 2024. Consent in Crisis: The Rapid Decline of the AI Data Commons. Data Provenance Initiative.

22 of 27

Widening inequality in data access

23 of 27

lobbying�open models = danger

24 of 27

lobbying�open models = danger

ignores efforts making open models safe

25 of 27

lobbying�open models = danger

ignores efforts making open models safe

brushes unfounded concerns under rug

26 of 27

PERSONAL OPINION�NOT MY EMPLOYER'S

I think we dodged a bullet 😊

27 of 27

Luca Soldaini��Twitter / X: @soldni

Bluesky: @soldaini.net�Email: lucas@allenai.org

Nathan Lambert��Twitter / X: @natolambert

Bluesky: @natolambert.bsky.social�Email: nathan@natolambert.com