Llama 2 opens commercial access

The release and the precise claim

On 18 July 2023, Meta published Llama 2, a family of large language models whose pretrained and chat-tuned weights could be downloaded and, for the first time in the Llama line, used inside many commercial products. The narrow, documented claim is specific. Meta released 7B, 13B and 70B parameter models (a 34B variant was trained but held back), together with a custom license, the Llama 2 Community License Agreement, that permitted commercial use subject to conditions. The accompanying paper, Llama 2: Open Foundation and Fine-Tuned Chat Models, reported that the models were pretrained on roughly two trillion tokens of publicly available data and that the chat variant, Llama 2-Chat, outperformed other open chat models on most of the benchmarks the authors ran.

The broad historical judgment is a separate matter, and the two are easily conflated. It is tempting to call this the moment open-source AI arrived, but that overstates both the license (which is not open source by the standard definition) and the novelty (downloadable models already existed). The more careful framing is that Llama 2 removed a specific commercial restriction on an unusually capable model family, and by doing so shifted a large share of applied language-model work off hosted APIs and onto weights that teams could run themselves. What follows keeps that documented change distinct from the mythology that accreted around it.

Llama 1 and the constraint of a research-only license

The predecessor makes the constraint concrete. In February 2023 Meta had released LLaMA, a set of foundation models from 7B to 65B parameters trained only on public data. The paper made a striking efficiency argument: its 13B model outperformed the 175B-parameter GPT-3 on most benchmarks, and its 65B model was competitive with Chinchilla-70B and PaLM-540B. The technical result was strong, but the weights were distributed under a noncommercial research license and gated by an application form, which meant a startup could study the model but not ship it.

That constraint collapsed almost immediately in practice. Within about a week the LLaMA weights leaked and circulated on public torrents, and a wave of derivative work (instruction-tuned variants, quantized builds that ran on laptops) followed. The lesson Meta drew, visible in the Llama 2 announcement's note that Llama 1 had drawn more than 100,000 access requests, was that demand for runnable weights was enormous and that a research-only posture neither satisfied it nor contained it. Llama 2 addressed that mismatch directly: the constraint it removed was legal and commercial, not primarily technical.

What "open weights" actually meant here

The phrase open weights describes a distribution model, not a legal status, and the distinction is the crux of this event. Under the Llama 2 Community License Agreement, anyone could download the parameters, fine-tune them, host them, and build commercial products on top, royalty-free. But the grant carried conditions that a genuinely open license would not. The most cited is Section 2: any company whose products had more than 700 million monthly active users in the calendar month before the release date had to request a separate license from Meta, a clause aimed squarely at the handful of hyperscale rivals.

Two further terms mattered. An Acceptable Use Policy prohibited enumerated categories of use, and the original license forbade using Llama 2 or its outputs to improve any other large language model that was not itself a Llama derivative, a clause that constrained competitors from bootstrapping rival systems on Llama-generated data. None of these conditions prevented the overwhelming majority of developers from building freely, which is why the release genuinely widened access. But they are conditions, and they are what separate this from software released under a permissive license such as Apache 2.0. The correct reading is that Llama 2 made a strong model commercially usable for almost everyone while reserving specific rights to Meta.

How the models were built and tuned

The base models are decoder-only Transformers, and their headline change from Llama 1 was scale and context. Pretraining used about two trillion tokens (up from the 1.0 to 1.4 trillion of Llama 1), and the context window doubled to 4,096 tokens. The larger 34B and 70B models adopted grouped-query attention, an inference-efficiency technique that shares key and value projections across groups of query heads, sitting between full multi-head attention and the more aggressive multi-query attention. That choice is a good example of the model's engineering character: the goal was not only quality but making a 70B model practical to serve.

The chat models are where most of the paper's methodological weight sits. Llama 2-Chat was produced through supervised fine-tuning followed by reinforcement learning from human feedback, the same broad recipe OpenAI documented for InstructGPT, but with distinctive Meta choices. The team trained two separate reward models, one for helpfulness and one for safety, so the two objectives could be optimized without one silently degrading the other, and combined rejection sampling with proximal policy optimization across several rounds. They also introduced a technique they called Ghost Attention (GAtt) to help the model respect a system instruction across many conversational turns. The unusual detail is how much of this pipeline the paper described in the open, including its safety and red-teaming process; the 34B model was withheld specifically because the authors said they had not finished red-teaming it in time.

The evidence: benchmarks and human evaluations

The paper's quantitative case rests on two kinds of evidence, and reading them precisely matters. On standard academic benchmarks, Llama 2 70B posted a score of roughly 68.9 on the five-shot Massive Multitask Language Understanding benchmark, a 57-subject multiple-choice test, and led contemporaneous open base models such as Falcon and MPT across most of the suite the authors reported. These are task-specific accuracy figures on defined test sets, not measures of general intelligence, and they say nothing about performance on tasks outside the benchmarks.

The second kind of evidence is human preference. Meta ran large-scale human evaluations comparing Llama 2-Chat against other systems on thousands of prompts, and reported that the 70B chat model was roughly comparable to the ChatGPT model of the period on helpfulness (a win rate near 36 percent with a substantial fraction of comparisons judged ties), while still trailing GPT-4. That is a meaningful result and an honestly hedged one: a preference win rate on a curated prompt set is a snapshot of relative helpfulness as judged by specific annotators, not a verdict that the open model matched the strongest proprietary system. The paper is careful about this, and a reader should be too.

The competitive landscape of mid-2023

Llama 2 did not arrive into an empty field, and its significance is clearer against its rivals. The dominant systems by capability were closed and API-only: OpenAI's GPT-4 and GPT-3.5, Anthropic's Claude, and Google's PaLM 2. On the open side, the Technology Innovation Institute had released Falcon models, later described in The Falcon Series of Open Language Models, under a genuinely permissive license, and MosaicML's MPT models were also downloadable. What Llama 2 offered that the closed leaders did not was self-hosting, and what it offered that the other open models mostly did not was a chat model tuned and evaluated to a level that made it a credible substitute for a hosted assistant in many applications.

The competitive picture also shifted quickly after July 2023, which cautions against treating Llama 2 as a permanent high-water mark. In September, a much smaller model, Mistral 7B, was released under Apache 2.0 and reported outperforming Llama 2 13B across every benchmark its authors tested, using both grouped-query attention and sliding-window attention. That a 7B model from a young European lab could beat a 13B Llama within two months underlines that Llama 2's importance was less about any single benchmark lead and more about the ecosystem posture it validated: a major company shipping capable weights with commercial permission.

Why the OSI said this is not open source

Because the release was so often described as open source, the Open Source Initiative published a direct rebuttal. In a July 2023 post, the OSI's Stefano Maffulli argued that the Llama 2 license fails the Open Source Definition on several counts, primarily its restrictions on certain fields of use through the Acceptable Use Policy and its differential treatment of very large licensees. The Open Source Definition requires, among other things, no discrimination against persons, groups, or fields of endeavor, and the OSI's position was that a license reserving rights against specific competitors and specific uses is, whatever its practical generosity, not open source.

This is not a pedantic quarrel, and it clarifies the event's real character. The distinction the OSI drew (between making resources available to most users under conditions, and granting the full freedoms of open source) is exactly the line this release straddles. The honest description, which Meta's own marketing sometimes blurred, is open weights with a commercial license, not open source. Getting that label right matters for anyone reasoning about legal risk, and it matters historically because the later fight over what open should mean in AI, including the emergence of terms like open weights, was partly provoked by this launch.

The downstream ecosystem it seeded

The clearest evidence of influence is what got built on Llama 2. Meta itself shipped Code Llama in August 2023, a family of code models built directly on Llama 2, and the broader community produced thousands of fine-tunes, quantized variants, and domain adaptations distributed through hubs like Hugging Face. Because the weights were commercially licensed, that activity was not confined to research demos: it fed real products, on-premise deployments in regulated industries, and the tooling ecosystem (efficient inference servers, parameter-efficient fine-tuning libraries) that now assumes downloadable weights as a starting point.

Meta's own trajectory shows how the posture evolved. Llama 3 followed in 2024, and Llama 3.1 in July 2024 added a 405B-parameter flagship, a 128K context window, and, notably, a license change that removed the earlier prohibition and explicitly allowed developers to use model outputs to improve other models. That later loosening is itself a reading of the Llama 2 experience: the constraints that drew the most criticism were the ones Meta relaxed. Tracing the line from Llama 1's research license, through Llama 2's conditional commercial grant, to Llama 3.1's more permissive terms is a better account of the influence than any single-release story.

The honest limits and caveats

Several caveats keep the event in proportion. First, open weights are not the same as open source, and they are also not the same as a reproducible model: Meta released the parameters but not the training data or the full pipeline needed to reproduce them from scratch, so the transparency is partial. Second, self-hosting a 70B model is not free. It demands GPUs, serving infrastructure, and operational expertise that a hosted API abstracts away, so the practical accessibility Llama 2 unlocked was greatest for teams that already had, or could rent, that capacity.

Third, the benchmark and preference numbers describe specific tasks and specific evaluations. Strong MMLU accuracy and a competitive helpfulness win rate against one version of ChatGPT do not establish general reliability, safety in deployment, or parity with the strongest closed models, and the Llama 2 paper is explicit about residual limitations including hallucination and the boundaries of its safety tuning. Finally, capable open weights carry a genuine dual-use tension: the same freedom that let a hospital run a model on its own hardware also lowered the barrier to misuse, which is precisely why the release shipped with an Acceptable Use Policy and a responsible-use guide rather than as an unconditional dump. None of these caveats negates the event; they define its actual shape.

How to place Llama 2 in AI history

The evidence supports treating 18 July 2023 as a licensing and distribution turning point rather than a purely technical one. The architecture was a well-executed decoder-only Transformer with sensible efficiency choices; the training followed the by-then-established scaling and RLHF playbook informed by work like compute-optimal scaling. What was new was the combination: a frontier-adjacent model family, from a company of Meta's scale, released with weights and a license permitting commercial use by nearly everyone. That combination is what moved the center of gravity in applied AI toward self-hosted models.

A useful way to navigate outward from this event is to hold three questions apart. What exactly was granted (downloadable weights under a conditional commercial license, not open source)? What did the release demonstrate rather than invent (that a major firm would ship capable weights, and that a large ecosystem would form when it did)? And which of its constraints later work removed (the 700-million-user clause and the output-use restriction, the latter relaxed by Llama 3.1)? Answered that way, Llama 2 reads not as the birth of open AI, a claim it cannot bear, but as the moment open-weight models became a mainstream, commercially viable alternative to the API, with all the freedom, responsibility, and unresolved definitional argument that shift carried with it.

References

• Touvron et al. (2023). Llama 2: Open Foundation and Fine-Tuned Chat Models.
• Meta (2023). Meta and Microsoft Introduce the Next Generation of Llama.
• Meta (2023). Llama 2 Community License Agreement.
• Touvron et al. (2023). LLaMA: Open and Efficient Foundation Language Models.
• Ainslie et al. (2023). GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints.
• Maffulli (2023). Meta's LLaMa license is not Open Source.
• Rozière et al. (2023). Code Llama: Open Foundation Models for Code.
• Jiang et al. (2023). Mistral 7B.
• Almazrouei et al. (2023). The Falcon Series of Open Language Models.
• Meta (2024). Introducing Llama 3.1: Our most capable models to date.
• Hoffmann et al. (2022). Training Compute-Optimal Large Language Models.
• Ouyang et al. (2022). Training language models to follow instructions with human feedback.
• Hendrycks et al. (2020). Measuring Massive Multitask Language Understanding.