Meta's Llama 4 journey has been one of the most dramatic stories in AI this year – from a triumphant launch to benchmark scandal, leading to a complete restructuring of Meta's AI organization.
Llama 4 Model Family
Scout (Available)
- Parameters: 17B active / 109B total (16 MoE experts)
- Context: 10 million tokens – unprecedented for open models
- Best for: Long workflows and massive data analysis
Maverick (Available)
- Parameters: 17B active (128 MoE experts)
- Best for: Coding, chatbots, and technical assistants
- Balance: Reasoning power vs response speed
Behemoth (Still Training)
- Outperforms GPT-4.5 and Claude Sonnet 3.7 on STEM benchmarks
- Teacher model for distilling knowledge to smaller models
- Release date TBD
The Benchmark Controversy
The most damaging revelation came from Meta's own chief AI scientist:
"We fudged a little bit" – Yann LeCun, admitting Meta used different model versions on different benchmarks to improve results.
Third-party researchers found their results didn't align with Meta's published benchmarks, causing:
- Loss of confidence among Meta leadership
- Internal frustration within the AI team
- CEO Mark Zuckerberg ordering an AI organization overhaul
Meta Superintelligence Labs (MSL)
In response, Zuckerberg announced the establishment of Meta Superintelligence Labs – a new division focused on:
- Building artificial superintelligence (ASI)
- Separate governance from existing AI research
- Direct reporting to Zuckerberg
- Competing directly with OpenAI and DeepMind
LlamaCon Announcements
At Meta's first-ever LlamaCon:
- Llama Guard 4: Safety filtering for AI outputs
- LlamaFirewall: Security layer for AI deployments
- Llama Prompt Guard 2: Protection against prompt injection
- Llama API: Limited preview combining open-source flexibility with closed-model features
Competitive Landscape
With Qwen 3 and DeepSeek V4 as open-weight competitors, Meta is betting on:




Comments (0)
Be the first to comment!