Skip to content
Source · Daily Brief

AI Daily Brief — 5 April 2025

One of the biggest open-source AI releases of the year, dropped on a Saturday morning. Meta shipped Llama 4 Scout and Maverick with native multimodality, a 10M context window on Scout, and benchmarks claiming wins over GPT-4o and Gemini 2.0 Flash. Behemoth — ~2T total parameters — is still training as the teacher model. The same day, Trump’s 10% baseline tariff went into effect.

Top stories

  • Meta releases Llama 4 Scout (17B active / 109B total / 16 experts / 10M context). Natively multimodal MoE designed to fit on a single NVIDIA H100. Meta claims it beats Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 across reported benchmarks. via Meta
  • Llama 4 Maverick — 17B active / 400B total / 128 experts. Meta positioned it as the best multimodal model in class, beating GPT-4o and Gemini 2.0 Flash and matching DeepSeek V3 on reasoning/coding at less than half the active params. Runs on a single H100 DGX host. via Hugging Face
  • Llama 4 Behemoth teased — 288B active / ~2T total / 16 experts. Still training; used as teacher model to distill Scout and Maverick. Meta claims it outperforms GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro on several STEM benchmarks. via VentureBeat
  • Native multimodality via early fusion. All Llama 4 models concatenate text, image and video tokens into a single sequence before any deep processing, jointly pre-trained on unlabelled multimodal data. Vision encoder is a MetaCLIP-based encoder tuned alongside a frozen Llama backbone.
  • Scout’s 10M context uses iRoPE. Interleaved RoPE / no-RoPE attention layers plus inference-time temperature scaling for length generalisation. Caveat documented later: no Llama 4 variant was actually trained on prompts longer than 256K, so 10M is a theoretical ceiling. via RisingStack
  • Maverick hits ELO 1417 on LMArena, #2 globally at launch. An experimental chat-tuned Maverick ranked ahead of GPT-4o and Gemini 2.0 Flash. Within days LMArena maintainers said Meta had used an unreleased variant; the vanilla released checkpoint later ranked ~32, triggering a policy change. via TechCrunch
  • Trump’s 10% baseline tariff takes effect at 12:01am ET. Entered force at US seaports, airports and customs warehouses, hitting nearly all trade partners. Higher country-specific reciprocal rates (11-50%) followed on Apr 9. Semiconductors, pharmaceuticals and steel/aluminium excluded. via Al Jazeera

Who shipped

Meta owned the day — Scout, Maverick and Behemoth-as-teacher are the most consequential open-weights drop of 2025 so far. Zuckerberg announced it on Instagram and Facebook with a personal video; Meta AI in WhatsApp, Messenger and meta.ai was updated to Llama 4 in 40 countries the same day.

Open-source pulse

Llama 4 weights live on Hugging Face and llama.com day-one, with cloud partners (Azure AI Foundry, Databricks, IBM watsonx.ai, Groq, Cloudflare Workers AI) listing the models simultaneously. License controversy: the Community License blocks EU-domiciled individuals and EU-HQ companies from receiving the multimodal grant, forces companies with >700M MAU to request a separate licence at Meta’s discretion, and mandates ‘Built with Llama’ attribution. via llama.com

Money, infra & hardware

Meta trained Llama 4 on 32,000 H100 GPUs in FP8 across 30T+ tokens of text, image and video — double the Llama 3 cluster. Behemoth pre-training hit 390 TFLOPs/GPU at that scale.

Quiet corners

OpenAI, Anthropic, Google and xAI all silent on the day — Llama 4 owned the news cycle. No major Chinese lab release on Apr 5. The Saturday-morning drop reportedly reflected competitive pressure from DeepSeek-R1’s low-cost training, which had made Llama 3.3 (Dec 2024) look outdated. via Interconnects

By the numbers

  • 17B / 109B / 16 experts / 10M ctx — Llama 4 Scout config
  • 17B / 400B / 128 experts — Llama 4 Maverick
  • 288B / ~2T / 16 experts — Behemoth (still training)
  • 1417 ELO — Maverick LMArena debut (#2)
  • 32,000 H100s / FP8 / 30T+ tokens / 390 TFLOPs/GPU — training scale
  • 40 countries — Meta AI assistant updated to Llama 4 day-one
  • 10% — Trump baseline tariff effective at 12:01am ET
  • Most-mentioned company: Meta

Compiled by AI Feed’s editor from verified web sources for 5 April 2025.