Meta has released Llama 4, the latest version of its open-source large language model, and the benchmarks put it within striking distance of GPT-4 Turbo on most standard evaluations. The release continues Meta's commitment to open-source AI and significantly raises the ceiling for what open-source models can do.
Benchmark Performance
Llama 4 70B (the mid-size version) scores 81% on MMLU, compared to GPT-4 Turbo's 86%. On coding benchmarks, it reaches 74% on HumanEval. The 405B parameter version (the largest model released) hits 84% on MMLU — genuinely competitive with closed-source frontier models at a fraction of the operational cost for organizations running their own infrastructure.
Meta's decision to release the 405B weights is the headline. Running a 405B model requires significant hardware, but research labs, universities, and large enterprises with existing GPU infrastructure can now access frontier-level capability without API costs.
What Open Source Means Here
Unlike GPT-4 or Claude, which are accessible only via API with usage costs and terms of service, Llama 4 can be downloaded, run locally, fine-tuned, and deployed however an organization needs. For enterprises with data privacy requirements, this is transformative — they can run a frontier-class AI without sending any data to a third-party API.
The open-source AI ecosystem that has built around previous Llama releases — fine-tunes, quantized models, tooling, deployment infrastructure — will rapidly adapt to Llama 4, further extending its practical value.
The Takeaway
Llama 4 is the most capable open-source AI model released to date, and it meaningfully narrows the gap with closed-source frontier models. For organizations that care about data privacy, control, or cost, this release is significant. The open-source AI ecosystem is no longer playing catch-up — it's arrived.