Meta AI July 2026 Updates: Llama 4, Assistant Changes, and More

Meta has had a busy July. Between rolling out improvements to Llama 4, expanding Meta AI across its apps, and releasing new open-source checkpoints, the company has continued its aggressive push to become a central player in AI infrastructure. Here's a clear breakdown of what changed this month and what it means for developers and regular users alike.

Llama 4 Scout Gets a Mid-Cycle Update

Meta's Llama 4 Scout — the mid-size model in the Llama 4 family optimized for speed and efficiency — received an updated checkpoint in July. The revision focuses on instruction following and code generation, two areas where the original Scout release lagged behind Mistral and Qwen at similar parameter counts.

Independent benchmarks on Hugging Face's Open LLM Leaderboard show the updated Scout scoring about 4 points higher on the MMLU-Pro coding subset and showing improved consistency on multi-turn conversations. For developers using Scout in agentic pipelines, the improvement in instruction following is particularly relevant — earlier versions would sometimes ignore constraint specifications in system prompts.

The updated Llama 4 Scout weights are available under Meta's Llama license at llama.meta.com, with both BF16 and quantized versions available.

Meta AI Assistant: New Features in WhatsApp and Instagram

On the consumer side, Meta has been quietly expanding the capabilities of its Meta AI Assistant across WhatsApp, Instagram, and Facebook. July's notable additions include:

• Voice mode for WhatsApp — Meta AI can now handle back-and-forth voice conversations in WhatsApp in 15 languages, up from 8. The voice quality has improved noticeably, with better handling of regional accents.
• Image editing in Instagram DMs — Users can now prompt Meta AI to edit photos shared in a direct message, including background removal, style transfer, and object replacement.
• Shopping recommendations — Meta AI in Facebook and Instagram now integrates with the Shops catalog to offer product recommendations based on conversation context.

These consumer features are powered by Meta's Llama 4 models running on Meta's internal inference infrastructure, not the publicly released open weights. Meta has been notably tight-lipped about which model variant powers consumer Meta AI, but performance is clearly above what the public Llama 4 Scout weights would suggest.

New Open-Source Code Model Release

One of July's more significant contributions to the open-source AI community was Meta's release of Code Llama 4 — a code-specialized variant fine-tuned from Llama 4 Maverick. Meta claims it matches or exceeds GPT-4o on HumanEval and achieves near-state-of-the-art results on SWE-bench Lite.

The model comes in three sizes (7B, 34B, and 70B) and supports 32 programming languages with strong emphasis on Python, JavaScript, TypeScript, and Rust. Fill-in-the-middle (FIM) capability has also been improved, which matters for IDE integration use cases.

For developers building AI coding tools on open-source foundations, Code Llama 4 is likely the most capable freely available option as of this writing.

Meta's AI Infrastructure Expansion

Meta quietly published details in July about a significant expansion of its AI training infrastructure. The company has brought additional GPU clusters online at its data centers in Iowa and Texas, with a combined capacity of over 600,000 H100-equivalent GPUs dedicated to AI workloads.

This matters because Meta's open-source strategy is partly funded by — and dependent on — its ability to train large models at low cost. More compute capacity means faster iteration on future Llama versions and more frequent open-source releases. AI chip demand continues to be the key constraint, and Meta's scale gives it negotiating power with Nvidia that most organizations don't have.

What Meta AI Still Doesn't Do Well

For all the progress, there are areas where Meta AI still lags. In direct comparison tests:

• Meta AI Assistant makes more factual errors on current events than Claude or Perplexity with search enabled
• Multi-document reasoning with Llama 4 public models still trails GPT-5 and Claude Opus 4 on complex tasks
• Safety filtering in the open-source models is intentionally minimal, which means deployers are responsible for their own guardrails

These aren't disqualifying weaknesses, but they're worth noting if you're evaluating Meta AI for production use cases.

Regulatory Considerations

Meta is navigating AI regulation carefully. The company released a transparency report this month covering its AI systems and their compliance with the EU AI Act's requirements for general-purpose AI models. The report outlines Meta's approach to capability evaluations, incident reporting, and model cards for Llama 4.

Meta has positioned open-source models as inherently more compliant with transparency requirements than closed models — an argument that has attracted both supporters and critics in the regulatory debate.

What to Expect in August and Q3

Meta has indicated that Llama 5 development is underway, though no release timeline has been given. A separate multimodal model focused on video understanding is also reportedly in late-stage development internally.

For the Meta AI Assistant, the company has previewed a "memory" feature similar to what OpenAI and Anthropic have rolled out — letting the assistant remember preferences and past context across sessions. That feature is expected to roll out broadly in Q3.

---

Looking to build with Meta's open models? The updated Llama 4 weights are available at llama.meta.com. For a broader comparison of AI models available today, see our guide to open-source AI models in 2026.