OSS frontier LLMs & multimodal arms race
Key Questions
What are Meta's plans for open-source AI models?
Meta plans a hybrid open-source strategy, releasing phased versions post-Llama4. Open-source editions will follow but not include all features of proprietary models.
What is Gemma4?
Gemma4 from Google DeepMind is a multimodal and agentic model with E4B parameters. It supports open-source local AI setups for agents and workflows.
What is Arcee-Trinity?
Arcee-Trinity is a 400B-parameter open-source LLM built by a 26-person U.S. startup on a $20M budget. It positions as a top OSS model from a tiny team.
Which cheap models achieved SOTA recently?
Qwen3.6, MiniMax, GLM, and Kimi are highlighted as cost-effective state-of-the-art models. They were among last week's biggest winners in open-source and cheaper categories.
What is Delangue's contribution to OSS agents?
Clement Delangue emphasizes building datasets for open-source frontier agents. His traces, along with Harrier, fuel replication efforts.
What is Mistral Small4?
Mistral Small4 is part of the tooling boom alongside CoPaw and Claw. It contributes to advancements in OSS multimodal and agentic capabilities.
How is Gemma being fine-tuned?
Tutorials exist for fine-tuning Gemma on TPU v5 using Kinetic, Keras, and JAX. This stack leverages hardware efficiently for OSS development.
What fuels the OSS frontier LLMs arms race?
Hybrid OSS plans from Meta, multimodal Gemma4, Arcee-Trinity's scale, cheap Chinese SOTA models, and tooling like CoPaw drive the competition and replication.
Meta hybrid OSS plans confirmed (phased post-Llama4); Gemma4 E4B multimodal/agentic; Arcee-Trinity 400B (tiny team top OSS); Qwen3.6/MiniMax/GLM/Kimi cheap SOTA; Delangue traces/Harrier fuel replication; Mistral Small4/CoPaw/Claw tooling boom.