Llama 3.1 Prelaunch and Release Timeline

Llama 3.1 Prelaunch and Release Timeline

Meta released Llama 3.1 on July 23, 2024. The family included 8B, 70B, and 405B text models with a 128K context length. Venture Step E026 was recorded shortly before that announcement, when the largest model's exact size, name, terms, artifacts, and real performance were still unsettled in the episode.

This timeline keeps anticipation and release evidence separate. It does not treat Meta's evaluations as independent proof, and it does not rewrite Dalton Anderson's pre-release discussion with facts that became available later.

timeline
    title Llama 3.1 claim states
    April 18, 2024 : Llama 3 8B and 70B released
    Before July 23, 2024 : E026 discusses an expected roughly 400B model
    July 23, 2024 : Meta announces Llama 3.1 8B, 70B, and 405B
    July 23, 2024 : Model card, license, paper, and artifacts establish release scope
    After release : Independent and workload-specific evaluation remains necessary

April 18, 2024: Llama 3 establishes the starting point

Meta released Llama 3 in 8B and 70B sizes. The official llama-models repository records the launch date, model sizes, context length, model card, license, and acceptable-use policy.

This matters because E026 compares the expected flagship with the smaller Llama 3 artifacts already available. The recording's comments about existing use are source-era observations, not a current product comparison.

Evidence state: released publisher documentation and artifacts.

Before July 23, 2024: E026 records anticipation

E026 refers to an expected 408-billion-parameter model and says a release should arrive later that week. Dalton discusses reports about download size, compute requirements, model weights, licensing, benchmark performance, and competition with closed models. He repeatedly signals uncertainty and says he wants direct access or independent review.

The episode supports a historical claim about what Dalton understood and expected at the time. It does not support presenting every pre-release detail as fact. The exact recording is preserved on YouTube.

Evidence state: dated first-person commentary drawing on mixed pre-release information.

July 23, 2024: Meta announces Llama 3.1

Meta's release announcement named three sizes: 8B, 70B, and 405B. It described a 128K context window, support for eight languages, tool-use improvements, distribution through Meta and partner platforms, and new safety components.

The announcement also characterized 405B as competitive with leading closed models and emphasized synthetic data generation and distillation. Those are publisher representations. They need to be labeled that way until independent evidence or a reader's own evaluation addresses the relevant task.

Evidence state: dated publisher announcement.

July 23, 2024: the model card defines the artifact

The official Llama 3.1 model card identifies pretrained and instruction-tuned text models in 8B, 70B, and 405B sizes. It records supported languages, a December 2023 knowledge cutoff, architecture details, intended use, safety guidance, evaluations, and a July 23 release date.

The model card is stronger evidence than a pre-release report for artifact identity and the publisher's stated scope. It is not a guarantee that a deployment will be accurate, safe, fast, affordable, or suitable for a particular use.

Evidence state: released publisher technical documentation.

July 23, 2024: access and legal terms become inspectable

The official repository provides access paths and identifies the governing Llama 3.1 Community License. The release also has an Acceptable Use Policy.

This resolved part of E026's uncertainty: official weights were made available. It did not make "open source" a sufficient legal or technical description. The rights, conditions, restrictions, source-code availability, training-data transparency, reproducibility, and practical ability to operate the model remain separate questions.

Evidence state: released legal terms and artifact-access documentation.

After release: performance becomes a workload question

Meta reported evaluations across more than 150 benchmark datasets and human comparisons. The model card contains additional capability and safety results. Those records are valuable, but they do not answer whether one variant works for a specific organization.

A deployment decision still requires an exact artifact, runtime, quantization, hardware plan, prompt or system design, representative dataset, baseline, quality thresholds, safety tests, latency, throughput, reliability, observability, and cost measurement. Episode 27's [[How to Evaluate an Open Weight Model Before Deployment]] turns that requirement into an adoption process.

Evidence state: publisher evaluation available; independent and use-specific evidence required.

What changed between expectation and release

The model was 405B rather than the 408B figure used in E026. Meta released official weights and documentation. The license became inspectable rather than speculative. The release included updated 8B and 70B variants, a longer context window, multilingual support, and a broader system story.

What did not become settled was equally important. Publisher benchmarks did not prove every workload. Download access did not erase infrastructure cost. A custom license did not become unrestricted permission. Safety components did not make a complete deployed system safe by default.

The clean editorial correction is not "the episode was wrong." It is "the episode recorded a pre-release state, and later artifacts resolved some claims while leaving others for evaluation."

For the method behind that correction, read [[How to Verify an AI Model Claim Before Release]]. For the access-language distinction, continue to [[Open Weight Is Not the Same as Open Source]]. E029 carries the discussion into training-data scaling and safety.

This timeline was developed with AI assistance from the recovered E026 recording and the linked official sources. Dalton Anderson remains the author. Meta is the primary source for its release and artifact claims. Historical, technical, license, source, accessibility, and founder review are required before publication. Publication is not authorized.