Llama——From Open-Source Experiment to the Most Deployed Open LLM, and Meta's Closed-Source Pivot

🌏 中文版

In February 2023, Meta released Llama 1—an open LLM family at 7B/13B/33B/65B. That decision changed the industry: open models went from "academic toy" to "enterprise infrastructure." By April 2026, Llama 4 Scout arrived with a 10M-token window and Maverick with a 128-expert MoE, both open multimodal benchmarks. But the same year Meta launched closed-source Muse Spark, and Llama 4 is likely the last major open Llama. This is the fifth family deep-dive in the "AI 模型家族" series, tracing Llama's complete evolution from Llama 1 to Llama 4 and its changed position in 2026.

For how to read the benchmark numbers cited here, see our AI model evaluation sources guide. This is part of the AI Model Landscape Overview series.

Family Evolution Timeline

Three years, nine milestones. Llama's through-line: from small to large, from text-only to multimodal, from open to (partly) closed. Llama 4 is the inflection—both the open peak and likely the last major open Llama.

Two Product Lines: Open Weights for Ecosystem, Closed Muse Spark for Flagship

To understand Llama's 2026 moves, split it into two parallel lines—like Qwen and Meta, open-first then recently closing:

Open-Weight Line (HuggingFace / Ollama / llama.cpp): Every generation from Llama 1 through Llama 4, fully downloadable, fine-tunable, self-hostable. This line owns ecosystem—Ollama, vLLM, llama.cpp all first-class; fine-tuning community's preferred base; nearly every local AI desktop app (LM Studio, Jan, GPT4All) runs Llama out of the box.

Closed-Flagship Line (Muse Spark): Only appeared April 2026, with Meta's announcement calling Llama 4 Maverick "our previous generation"—marking the flagship's shift from open to closed. This line never releases weights; future updates are Meta-only.

The turning point: Meta (ex-Facebook) has, since establishing FAIR in 2013, used "open source" as core strategy—not by selling model API (unlike OpenAI and Anthropic), but by enhancing social products and collecting hosting fees via Azure/AWS. Llama is the vehicle. Muse Spark's closure is a strategic pivot (against Claude, GPT closure pressure and rising training costs), not a return to roots—but it does leave "will Llama stay open?" as an open question.

Architecture: Three Key Designs of Llama 4

MoE: Only 17B Active Per Step

Llama 4 is Meta's first fully Mixture-of-Experts family. Maverick example:

Total params:    400B
Active params:   17B (activated per inference step)
Experts:         128 routed + 1 shared
Sparsity:        ~24:1

Each token hits the shared expert plus one of 128 routed experts. All parameters sit in memory, but only a fraction activates per step—letting a 400B-class model serve on a single H100 host while staying frontier quality. Scout is 17B active / 109B total, 16 experts, designed to fit a single H100 quantized to Int4.

Native Multimodality: Early Fusion

Llama 4 uses early fusion—pre-training data mixes text and images from the start, not a vision adapter bolted onto a text model after training. Scout and Maverick natively understand images, not just "a text model with a vision add-on."

10M Context: Scout's Killer Feature

Scout's 10M-token context is the longest of any model (open or closed). A single request can hold an entire codebase, a whole book, or hours of transcript. Caveat: 10M is the documented upper bound; actual usable context depends on hardware and serving stack—after Int4 quantization Scout fits a single H100, but full 10M context still needs more KV-cache memory.

Llama 4: Scout, Maverick, Behemoth—How to Choose

2026's three naming slots serve distinct roles:

Pricing (via third-party hosts; Meta offers no first-party API):

Scout's $0.08 input is the cheapest of all models; paired with 10M context, it's the value king for long-document work.

License Trap: Open, But Not As Open As You Think

Llama 4 uses the Llama 4 Community License—not OSI-certified open source. Key restrictions:

• Products with >700M MAU using Llama 4 must seek separate authorization from Meta (approval at Meta's discretion)
• Must display "Built with Llama" on the product
• Has an Acceptable Use Policy

For most enterprises, this is near-open. For mega-platforms (Twitter, Snapchat, etc.), it's a real legal constraint. Compared to Apache 2.0 (most Qwen models) or MIT (GLM-5, DeepSeek), the permissiveness gap is clear—if your deployment depends on license certainty, "downloadable weights" ≠ "freely commercial."

Performance Position

Direct comparison:

Llama 4 still owns long-context and cost, but trails frontier by ~3pp on engineering benchmarks. On the "openness" spectrum it sits in the "downloadable but conditional" middle—more restrictive than true open (Apache/MIT), more free than fully closed.

Sub-lines & Ecosystem: A Table of All Llama Models

Llama runs multiple sub-lines beyond the general main line:

Two trends in this table:

Capability consolidated into the main line. Code Llama independent iteration ended—same playbook as Qwen and DeepSeek consolidating sub-lines. Once general models' specialist capability was strong enough, maintaining standalone sub-lines no longer paid. Vision took a different path: not cut, but turned into native capability via early fusion.

Ecosystem breadth is the hidden moat. Llama's real moat isn't a single model but ecosystem. Llama.cpp, Ollama, vLLM, Hugging Face Transformers all first-class; fine-tuning community's preferred base; most local AI desktop apps run Llama by default. That "runs everywhere" penetration makes Llama the de facto standard for open inference.

Final reminder: model suffixes (Scout/Maverick/Behemoth) don't guarantee downloadable weights (Behemoth never released). Check license and release status for self-host, not the suffix.

Position Against Competitors

Placing Llama 4 in the 2026 open landscape:

• vs Qwen: Qwen3.8-Max (2.4T) total and active params dwarf Maverick, and most Qwen models ship Apache 2.0 for better license certainty. Llama's edge is ecosystem maturity—still ahead on local inference framework support
• vs DeepSeek V4: DeepSeek's MLA architecture is cheaper, pricing more aggressive ($0.28/$0.42); Llama 4 exclusive on 10M context
• vs Kimi K3 (2.8T open): K3 total params larger, but Llama 4 Scout's 10M context remains the longest
• vs Closed (Claude / GPT / Gemini): Llama 4 trails ~3pp on SWE-bench and other engineering benchmarks, and Meta has pivoted to closed Muse Spark with uncertain future updates
• vs Mistral / GLM: Llama's ecosystem penetration still leads, but Mistral's license is cleaner (Apache/Modified MIT) and GLM pursues domestic compliance

What This Means for Agent Developers

• Ultra-long document analysis → Llama 4 Scout: 10M context + $0.08/MTok, unbeatable value right now
• Cost-sensitive enterprise deployment → Scout / Maverick self-hostable, data never leaves your infra—fits healthcare, finance data-sovereignty needs
• Multilingual scenarios → Llama 4 supports 200 languages (Arabic, Spanish, German, Hindi, etc.)
• Local / edge inference → Llama.cpp + quantization lets Llama run on laptops, even phones
• Need license certainty → Community License's 700M MAU clause; mega-platforms should use MIT/Apache models (GLM, DeepSeek, most Qwen)
• Need frontier coding agent → Llama 4 trails Claude and DeepSeek—pick those instead
• Citing benchmarks → Llama's naming matrix (gen × Scout/Maverick/Behemoth) is easy to confuse with unreleased Behemoth. Always write the full model name plus date, otherwise you're comparing different events

Overall

Llama's story is "how open source changed an industry." Meta's Llama family proved open models could hit commercial quality, spawning countless Llama-based products worldwide. Llama 4 Scout's 10M context and Maverick's 128-expert MoE remain technical benchmarks for open models; ecosystem penetration (Llama.cpp as the local inference byword) is unmatched.

But 2026's turn is also clear—Meta launched closed Muse Spark, calling Llama 4 Maverick "our previous generation." Two signals: Meta has moved its flagship from open to closed; Llama 4's future updates are uncertain. For agent developers, Llama 4 still owns ultra-long context and cost, but Community License (not OSI open) and Meta's strategic pivot are adoption risks. On the "openness" spectrum, Llama 4 sits "downloadable but conditional"—one more constraint than true open (Apache/MIT), one more freedom than fully closed.

References

• Llama 4 — Meta AI Official Blog
• Llama 4 Scout and Maverick on Hugging Face
• Llama 4 Model Card
• Meta Llama Pricing Guide 2026 — AI Cost Check
• Llama 4, reviewed — benchr
• Meta Superintelligence Labs / Muse Spark
• Llama.cpp — Local Inference Framework
• AI Model Evaluation Sources Guide — this site
• AI Model Landscape Overview — this site