AI Daily Brief — 6 April 2025
The Sunday after Llama 4. Within 24 hours of Meta’s Saturday drop, developers spotted that the LMArena-topping Maverick variant didn’t match the public weights — and the benchmark-manipulation accusations began. Trump defended his tariff schedule from Air Force One ahead of Wednesday’s reciprocal effective date.
Top stories
- Llama 4 LMArena controversy erupts. Developers spotted that the Llama-4-Maverick-03-26-Experimental variant scoring #2 on LMArena (ELO 1417) was a custom human-preference-tuned version not matching the public weights — sparking accusations of benchmark manipulation. The public Maverick weights would later rank #32. LMArena said Meta’s interpretation of policy ‘did not match what we expect.’ via The Register
- Slashdot frames Llama 4 as ‘milestone for open source.’ Zuckerberg’s claim that ‘for the first time, the best small, mid-size, and potentially soon frontier LLMs will be open source’ echoed through weekend discourse. Scout (17B active / 109B total, 10M context, single-GPU) and Maverick (17B active / ~400B total, 128 experts) live on Llama.com and Hugging Face; Behemoth (288B active / ~2T total) still training. via Slashdot
- Hugging Face ships day-one Llama 4 integration. Transformers v4.51.0 added native Scout & Maverick support including multimodal inference, with TGI and vLLM teams contributing serving support. Models hosted at meta-llama/Llama-4-Scout-17B-16E and Llama-4-Maverick-17B-128E on the Hub. via Hugging Face
- Trump posts tariff defiance from Air Force One. En route to Joint Base Andrews, Trump posted on Truth Social defending his ‘reciprocal’ tariffs ahead of their Apr 9 effective date, calling out China’s 34% retaliatory announcement and telling abusing countries not to retaliate. The weekend set up the Apr 7-9 stock-market crash that would hit AI chip stocks particularly hard.
Who shipped
Nothing new — Sunday after Llama 4. Meta kept absorbing the Llama 4 reception while the LMArena story metastasized.
Open-source pulse
Llama 4 weights live on Hugging Face and llama.com. Day-one community tooling — quantizations, inference engines, LangChain integrations — appeared rapidly through the weekend.
Quiet corners
No new model releases dated Apr 6. OpenAI, Anthropic, Google, xAI, DeepSeek and Qwen all silent on the day.
By the numbers
- ~24 hours — from Llama 4 launch to the first benchmark-manipulation allegations
- v4.51.0 — Hugging Face Transformers release with day-one Llama 4 support
- 3 days — until the higher Apr 9 country-specific reciprocal tariffs kick in
- Most-mentioned company: Meta
Compiled by AI Feed’s editor from verified web sources for 6 April 2025.