Meta’s new generation Llama 4 series open-source models are here! According to official test data, Llama 4 outperforms competitors like Gemma 3, Mistral, GPT-4o, and Gemini 2.0 Flash in many areas. Also, unlike previous versions, the three models launched this time have very large parameter counts, with Llama 4 Behemoth having 288B parameters. Most users probably won’t be able to install and use it locally, but that’s okay—there are free options available online to try it out.
Meta’s new generation Llama 4 series models officially debut.
This time, the Llama 4 series models come in three versions.
- Llama 4 Scout: a model with 17B active parameters, 109B total parameters, and 16 experts, is the most powerful multimodal model in its class, going head-to-head with Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1.
- Llama 4 Maverick: a model with 17B active parameters, 400B total parameters, and 128 experts, and the best multimodal model in its class, benchmarked against GPT-4o and Gemini 2.0 Flash.
- Llama 4 Behemoth, a model with 288B active parameters, 2T total parameters, and 16 experts, is currently Meta’s most powerful model, going head-to-head with GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro.
The Llama 4 model uses a Mixture of Experts (MoE) architecture, which is why each version is divided into active parameters and total parameters. The difference between the two is that the former refers to the number of parameters actually activated and involved in computation when processing a specific input, while the latter refers to the sum of all parameters in the model.
First, let’s look at Llama 4 Scout. Meta says Llama 4 Scout is a general-purpose model with supported context length greatly increased from 128K in the previous generation Llama 3 to 10 million tokens, meaning whether it’s multi-document summarization, parsing large amounts of activity to achieve personalization tasks, or reasoning over massive codebases, it can handle it all.
Below is the test data. From the LiveCodeBench scores, you can see that Llama 4 Scout is still slightly behind Llama 3.3 70B, but it beats Llama 3.1 405B, Gemma 3 27B, and Gemini 2.0 Flash-Lite. As for the other areas, such as image reasoning/understanding, reasoning and knowledge, and long-context comprehension, Llama 4 Scout comes out on top in all of them.
Llama 4 Maverick is a model suitable for image understanding and creative writing. Compared to Llama 3.3 70B, it offers lower cost and higher quality, and is the best-in-class multimodal model across coding, reasoning, multilingual, long-context, and image benchmarks.
Llama 4 Maverick surpasses GPT-4o and Gemini 2.0 in nearly every benchmark, and in coding and reasoning, it’s on par with the much larger DeepSeek v3.1.
Lastly, Llama 4 Behemoth, still in preview, is touted by Meta as the most advanced teacher model in its class, delivering impressive performance on math, multilingual processing, and image benchmarks.
Benchmark tests show that Llama 4 Behemoth outperformed Claude Sonnet 3.7, Gemini 2.0 Pro, and GPT-4.5 in every category:
Llama 4 Maverick and Llama 4 Scout models are now available for download on llama.com and Hugging Face. For those with H100s, Meta’s Llama 4 The introduction mentions that it can run on a device with a single NVIDIA H100 GPU.
Online part,Meta AI web versionLlama 4 has already been made available, but it hasn’t launched in Taiwan and most other countries yet. If you want to try it out, you can use OpenRouter。
How to use the Llama 4 model for free?
Currently, there are several methods on the internet,OpenRouter It’s the simplest method I’ve tried.
After clicking the link above to go to the OpenRouter website, you need to log in to your account. If you don’t have one, just register for free:
Supports multiple registration methods such as GitHub, Google, Email:
After registering, go to Chat and click the Add model function at the top right:
You can find Llama 4 Maverick (free) and Llama 4 Scout (free) models in the menu:
When you enter the content you want to chat about, an error pops up:
Click the link in the error:
Turn on the Model Training option:
Returning to the chat window makes conversation work normally again:
However, note that OpenRouter’s Llama 4 Maverick (free) is provided by Chutes AI, while Llama 4 Scout (free) is from DeepInfra, not Meta itself, so it’s uncertain whether the models themselves are quantized.
I also asked the AI whether they were the Llama 4 model, and the answer was no, it’s Llama 3, so everyone can verify this point on their own.
In OpenRouter’s comparison table, Llama 4 Maverick (free) is also fp8, with tokens reaching 256K, which is higher than the paid version:
Source: KOCPC Chinese