What happened
Meta released Llama 4.2 Code 70B on July 24 as an open-weight coding model. It features a 1 million token context window and improved agentic tool-use reliability compared with Llama 4.1 Code.
Meta said the model matches Gemini 3.5 Pro on several internal coding benchmarks while running on two H100 GPUs with quantization. Weights and a fine-tuning recipe are on Hugging Face under the Llama license.
Why it matters
Open-weight coding models are closing the gap with closed frontier systems on the tasks teams care about most. A 1M context window also lets agents hold entire repositories in memory, which changes how multi-file refactors are architected.
Meta's release pressure also gives enterprise teams leverage when negotiating closed API pricing.
MintedBrain take
Open weights do not mean free inference. Budget for GPU hosting and evaluate quantized performance on your stack before committing. The 1M context is compelling for large repos, but test latency on real prompts first.
Discussion
Sign in to comment. Your account must be at least 1 day old.