Llama 4: Forecast, Release, and Successor Record

Llama 4 Forecast, Release, and Successor Record

E014 predicted that Llama 4 would focus on coordinating agents and autonomous task management. The reviewed pre-release sources do not show Meta making that commitment.

The actual Llama 4 release emphasized a different set of model properties. Later Meta AI products moved toward action, but Meta said those 2026 features were powered by Muse Spark 1.1.

timeline
    title Forecast and release record
    2024-04 : E014 forecasts coordinating assistants
    2025-04-05 : Llama 4 Scout and Maverick released
    2025-04-29 : Llama API limited preview announced
    2026-04 : Muse Spark powers rebuilt Meta AI
    2026-07 : Muse Spark 1.1 supports product actions

What E014 forecast

The recovered source imagined a personal assistant, called Alfred, that could understand a user's needs, contact specialized agents, schedule work, order goods, and coordinate a day.

It attached that direction to Llama 4. The source was making a forecast, not reporting a documented model roadmap.

That distinction should remain visible even when a later product moves in a similar direction.

What Meta released as Llama 4

Meta released Llama 4 Scout and Maverick on April 5, 2025.

The company described both as natively multimodal, open-weight mixture-of-experts models. Scout had 17 billion active parameters across 16 experts and a stated 10-million-token context capability under Meta's conditions. Maverick had 17 billion active parameters across 128 experts and 400 billion total parameters.

Meta also discussed Behemoth as a teacher model still in training. A previewed or training model is not a released model.

The benchmark comparisons on the launch page are vendor claims tied to disclosed tests and versions. They do not establish application performance, cost, safety, or fit for every workload.

Model access

Meta made Scout and Maverick available for download under its license and through partners. The current Llama getting-started page continues to distinguish the models, access routes, hosting, and supporting tools.

"Open-weight" should not be reduced to "open source" without examining the license and use conditions. Access to weights also does not remove compute, deployment, security, evaluation, data, and maintenance obligations.

LlamaCon and the API

At LlamaCon later in April 2025, Meta announced a limited-preview Llama API, SDKs, evaluation, fine-tuning, and protection tools.

That was developer infrastructure. It did not make every future Meta AI product feature part of the open-weight Llama release.

The Meta AI app launched with Llama 4, but the assistant product included its own interface, memory, voice, feed, account, and glasses integration.

The successor product model

In April 2026 Meta introduced Muse Spark as a product-focused model powering Meta AI.

In July, Meta described Muse Spark 1.1 enabling planning, email and calendar connections, slide creation, research, and task handling in the app and on meta.ai.

Those features resemble the class of coordinating assistant E014 imagined. The model name and product implementation differ from the forecast.

The correct statement is that later Meta AI product behavior moved toward some forecasted actions. It is not that Llama 4 arrived exactly as predicted.

Why the distinction matters

A forecast can be directionally useful while being wrong about the responsible model, date, architecture, access, or product.

If later similarity is treated as complete confirmation, forecasting becomes impossible to score. The record should preserve the original claim before examining the outcome.

This is the same discipline used in decision journals: define the forecast, deadline, resolution, evidence, and possible partial outcomes before the result.

How to evaluate Llama 4 now

For a model decision, verify the exact variant, weights, license, context conditions, input types, benchmark version, hardware, precision, latency, cost, safety tools, data path, support, and lifecycle.

For a Meta AI decision, verify the assistant surface, account, model, memory, connectors, action authority, confirmation, logs, correction, rollback, and policy.

Do not use a model benchmark to prove an agent workflow. Do not use a product demo to prove the open-weight model can reproduce the same system.

The maintained [[Meta Llama 4 Model Profile]] owns current model history if it remains actively maintained. This E014 page owns the original forecast, the actual 2025 release, and the later successor-product distinction.

A scoring rule for the forecast

The original forecast should be scored in parts. Meta did release a Llama 4 family, but its launch centered multimodality, long context, and mixture-of-experts models. Meta later shipped coordinating and connected-app behavior, but it attributed those features to Muse Spark 1.1 inside Meta AI.

That is partial directional similarity with material differences in model, release, timing, and product. Preserving those differences makes the record useful for future forecasts instead of rewarding any later event that resembles the original idea.

AI assisted with research organization, structure, drafting, and validation. Dalton Anderson remains the attributed author and final editorial authority. The transcript and linked public sources control factual claims. Publication remains unauthorized.