SwiftRecapSwiftRecap
Back to Archive
January 15 - January 21, 20265 min read

Open Source AI Models Close the Gap

AI & MLDev ToolsStartups

A watershed week for open-source AI. DeepSeek's R2 model matches GPT-4 performance at 1/10th the cost, while Meta's Llama 4 signals a new era of competitive open foundation models.

Top Stories

1

DeepSeek R2 Matches GPT-4 at $0.002 per 1M Tokens

The Chinese AI lab released R2, an open-weight model that achieves GPT-4-level performance on most benchmarks while being dramatically cheaper to run.

Why it matters: Price pressure on closed models will force OpenAI and Anthropic to cut costs or lose market share.

Read more5 min read
2

Meta Announces Llama 4 with Multimodal Reasoning

Llama 4 features native image understanding, video analysis, and improved reasoning. Released under a commercial-friendly license for companies under 700M users.

Why it matters: Gives startups and enterprises a viable alternative to proprietary models without usage restrictions.

Read more4 min read
3

Mistral Releases Code Model Beating GPT-4 on SWE-Bench

Mistral's new code-specific model achieves 48% on SWE-Bench, surpassing GPT-4's 43%. Optimized for software engineering tasks.

Why it matters: Specialized models are beating general models in their domains - fragmentation incoming.

Read more4 min read
4

Stability AI Releases Video Model Stable Video 2.0

Open-weight video generation model produces 4K quality video up to 60 seconds. Competitive with Runway and Pika.

Why it matters: Democratizes high-quality video generation for creators and small studios.

Read more3 min read
5

Hugging Face Smashes Fundraising Record with $1B Round

The AI model hub raised $1B at a $12B valuation, becoming the most valuable open-source AI company. Plans to build 'GitHub for AI models'.

Why it matters: Validation that open models and the ecosystem around them are a massive business opportunity.

Read more3 min read

Quick Hits

Google Open Sources Gemma 3: 27B Parameter Beast

Compact but powerful model for edge devices...

AI2 Releases OLMo 2: Fully Open Training Pipeline

Complete transparency in training data and methods...

Cohere Command R+ Goes Free for Research

Enterprise-focused model available for academic use...