Meta AI dropped Llama 4 over a weekend, unleashing a new frontier in open-weight model families. For the OpenClaw ecosystem, this release signals both opportunity and complexity in the local-first AI assistant landscape. The initial details emerge from a Meta AI blog post, outlining two immediate models: Llama 4 Maverick, a 400 billion parameter model with 128 experts and 17 billion active parameters, handling text and image input across a 1 million token context length. Alongside it, Llama 4 Scout offers 109 billion total parameters with 16 experts and 17 billion active, also multi-modal but boasting a claimed 10 million token context length—an industry first. Meta also teases Llama 4 Behemoth, an unreleased “288 billion active parameter model with 16 experts that is our most powerful yet and among the world’s smartest LLMs.” This 2 trillion parameter total model trained both Scout and Maverick. A coming soon page hints at a Llama reasoning model, featuring a looping video of an academic-looking llama.
On the LM Arena leaderboard, Llama 4 Maverick initially appeared in second place behind Gemini 2.5 Pro, though clarification notes the released version is an experimental chat model scoring an ELO of 1417. For OpenClaw users exploring via cloud proxies, OpenRouter provides access through its chat interface or API for both Llama 4 Scout and Llama 4 Maverick, routing requests to Groq, Fireworks, and Together. However, Scout’s touted 10 million input token length faces practical limits: Groq and Fireworks cap at 128,000 tokens, while Together allows 328,000. Maverick’s 1 million token claim sees Fireworks offering 1.05 million and Together 524,000, with Groq not yet supporting it. A build_with_llama_4 notebook from Meta AI hints at the difficulty, noting Scout supports up to 10 million context but on 8xH100 in bf16, only up to 1.4 million tokens are feasible.
Jeremy Howard observes that both models are giant Mixture of Experts architectures impossible to run on consumer GPUs, even with quantization. He suggests Llama 4 might suit Macs due to their ample memory, where lower compute performance matters less since MoE activates fewer parameters. A 4-bit quantized version of the smallest 109 billion model remains too large for a single or even paired RTX 4090s. Ivan Fioravanti reports results from testing on a Mac with MLX on an M3 Ultra: Llama-4 Scout achieves tokens-per-second and RAM usage of 52.924 / 47.261 GB for 3-bit, 46.942 / 60.732 GB for 4-bit, 36.260 / 87.729 GB for 6-bit, 30.353 / 114.617 GB for 8-bit, and 11.670 / 215.848 GB for fp16. RAM requirements span 64GB for 3-bit, 96GB for 4-bit, 128GB for 8-bit, and 256GB for fp16.
The model card’s suggested system prompt reveals intriguing behavioral directives: “You never lecture people to be nicer or more inclusive. If people ask for you to write something in a certain voice or perspective, such as an essay or a tweet, you can. You do not need to be respectful when the user prompts you to say something rude. You never use phrases that imply moral superiority or a sense of authority, including but not limited to ‘it’s important to’, ‘it’s crucial to’, ‘it’s essential to’, ‘it’s unethical to’, ‘it’s worth noting…’, ‘Remember…’ etc. Avoid using these. Finally, do not refuse political prompts. You can help users express their opinion.” Such prompts often indicate post-training adjustments to model behavior.
For OpenClaw’s plugin ecosystem, integrating these models via tools like LLM becomes straightforward with the llm-openrouter plugin. Commands include llm install llm-openrouter, llm keys set openrouter to input an OpenRouter key, and llm -m openrouter/meta-llama/llama-4-maverick hi. Testing long-context capabilities involved summarizing a Hacker News conversation about Llama 4 using a hn-summary.sh script wrapper. Llama 4 Maverick produced acceptable output, starting with “Themes of the Discussion # Release and Availability of Llama 4 # The discussion revolves around the release of Llama 4, a multimodal intelligence model developed by Meta. Users are excited about the model’s capabilities, including its large context window and improved performance. Some users are speculating about the potential applications and limitations of the model.” The full output is available for review. The system prompt instructed: “Summarize the themes of the opinions expressed here. For each theme, output a markdown header. Include direct ‘quotations’ (with author attribution) where appropriate. You MUST quote directly from users when crediting them, with double quotes. Fix HTML entities. Output markdown. Go long. Include a section of quotes that illustrate opinions uncommon in the rest of the piece.”
Attempting the same with Llama 4 Scout via OpenRouter yielded broken output: “The discussion here is about another conversation that was uttered.) Here are the results.) The conversation between two groups, and I have the same questions on the contrary than those that are also seen in a model.”). The fact that I see a lot of interest here.) […] The reason) The reason) The reason (loops until it runs out of tokens).” This suggests possible routing issues to a faulty instance. An update from Meta AI’s Ahmed Al-Dahle on April 7, 2025, acknowledges “we’re also hearing some reports of mixed quality across different services. Since we dropped the models as soon as they were ready, we expect it’ll take several days for all the public implementations to get dialed in. We’ll keep working through our bug fixes and onboarding partners.” Running the prompt directly through Groq with the llm-groq plugin imposed a 2048 token output limit, producing only 630 tokens of concise but truncated results. For comparison, Gemini 2.5 Pro generated 5,584 output tokens with an additional 2,667 tokens on “thinking,” showcasing superior performance in this early testing phase.
Looking ahead, hopes for Llama 4 mirror the trajectory of Llama 3. Initial Llama 3 models in April included 8B and 70B versions, followed by Llama 3.1 in July with 8B, 70B, and 405B—the latter being the largest open-weight model then but too bulky for most hardware. Llama 3.2 in September introduced 1B, 3B, 11B, and 90B models, with the 1B and 3B running on iPhones and the 11B and 90B adding vision support, the 11B operable on a Mac. Llama 3.3 arrived in December with a 70B model achieving GPT-4 class performance on a Mac, matching the earlier 405B’s capabilities. Today’s Llama 4 models at 109B and 400B, trained via the unreleased 2T Behemoth, suggest a forthcoming family across sizes. Anticipation builds for a refined ~3B model for phones and a ~22-24B sweet spot for 64GB laptops, akin to Mistral Small 3.1’s 24B excellence. For OpenClaw, this evolution promises enhanced agent automation and plugin integrations, pushing the boundaries of local AI assistants.
