Meta Releases Llama 4 Scout: Open-Source Model Rivaling GPT-4
Meta has released Llama 4 Scout, the latest iteration of its open-source large language model family. The 70B parameter model achieves scores within 3% of GPT-4 on MMLU, HumanEval, and GSM8K benchmarks, making it the most capable fully open-weight model available. Llama 4 Scout introduces several architectural innovations including grouped-query attention with 128 KV heads, […]
Meta has released Llama 4 Scout, the latest iteration of its open-source large language model family. The 70B parameter model achieves scores within 3% of GPT-4 on MMLU, HumanEval, and GSM8K benchmarks, making it the most capable fully open-weight model available.
Llama 4 Scout introduces several architectural innovations including grouped-query attention with 128 KV heads, rotary positional embeddings extended to 128K context, and a novel training recipe that combines pre-training on 15 trillion tokens with iterative preference optimization.
Meta is releasing the model under the Llama 4 Community License, which permits commercial use with minimal restrictions. The weights are available on Hugging Face, and inference code has been integrated into all major serving frameworks including vLLM, TGI, and Ollama.
Related Articles
Claude 5 Launches with Enhanced Reasoning and 1M Context Window
Anthropic has officially launched Claude 5, marking a major leap in AI capabilities. The new model introduces an enhanced reasoning
GPT-5 Turbo Achieves New SOTA on Multimodal Benchmarks
OpenAI has unveiled GPT-5 Turbo, its most capable multimodal model to date. The model achieves new state-of-the-art results on MMMU
Apple Intelligence 2.0 Brings On-Device AI Agents to iPhone
Apple has unveiled Apple Intelligence 2.0, a significant expansion of its on-device AI capabilities that brings autonomous agent functionality to