Today we release Antidoom, an open-source method that removes a common failure mode in reasoning models: the doom loop.
Doom-loop rates before and after, with eval scores up across the board:
> Early LFM2.5-2.6B checkpoint: 10.2% → 1.4%
> Qwen3.5-4B: 22.9% → 1% (greedy sampling)
🧵
Just dropped: The Gemma 4 Technical Report!
Dive into the decisions behind our latest open-weight multimodal models, Gemma 4 (E2B–31B). Discover details on how the Gemma team approaches architecture, efficiency and responsibility. The report also covers the recent encoder-free 12B variant, quantization-aware training and multi-token prediction (MTP) drafters.
🚨 New Meta Llama 5 leak just surfaced:
Meta appears to be prepping its next massive frontier model to take on OpenAI directly:
• Internal codename: Watermelon (believed to be Llama 5)
• Currently in training on roughly 10x the compute of April's Muse Spark release
• Performance is already matching GPT-5.5 on key internal benchmarks
• Expect major agentic and coding gains coming to Meta AI alongside a new developer API
• A dedicated Opus-level coding model is also coming "pretty soon," per Meta superintelligence lead Alexandr Wang
Are you looking forward to this model being released?