The version of Meta's Llama 2026 roadmap that nobody else writes: there is no Llama 5, and Behemoth is a teacher model.

The 13 months between Llama 4 Maverick and Muse Spark, in order

Most of the pages that try to answer "when does Llama 5 come out" were written before April 8, 2026 and forecast forward from rumour. None of that matters now. The timeline below is the actual list of shipped events from the last frontier Llama release to today, drawn from Meta's own blog, Meta's About newsroom, Wikipedia's Llama and Muse Spark entries, and Wall Street Journal reporting on the Behemoth slip. If a date is missing from this list, no public announcement landed for it.

Two practical readings from the timeline. First, the question "when does Llama 5 ship" resolves negatively. As of the cutoff date on this page, there is no Llama 5. The internal codename that some August 2025 reporting referred to as Llama 4.X or Llama 4.5 also never appeared under those labels. The release-date table for 2026 has exactly one entry from Meta's frontier line, and it is Muse Spark. Second, the question "when does Behemoth ship" resolves as "in practice, it doesn't." Behemoth was announced as still in training on April 5, 2025. The WSJ slip pushed it from early summer 2025 to fall 2025. After fall, the public communications stop. The model's job in 2026 is to teach the smaller herd members via codistillation, not to be downloaded.

Why this question really gets asked: where does my local Llama plug into the agent

People who type this query into Google are rarely looking for a corporate strategy memo. They are usually trying to decide what model to download next, whether to wait for a bigger Llama, and how it slots into whatever they actually run. For a real worked example, here is how that decision lands inside Fazm, a shipping Mac computer-use agent. Fazm's default backend is Claude through the Anthropic Agent Protocol. The product also supports routing every chat turn through a custom Anthropic-compatible base URL. That single base URL is the seam where any local Llama lives, regardless of whether it is Llama 3 8B on Ollama, Llama 3 70B quantised on llama.cpp, or Llama 4 Scout via a transformer-serving stack.

The wiring is concentrated in two files. The settings field is in Desktop/Sources/MainWindow/Pages/SettingsPage.swift. Around line 962 the label reads "Custom API Endpoint." Line 984 is a TextField bound to an @AppStorage key named "customApiEndpoint." The agent subprocess reads that key back at startup, in Desktop/Sources/Chat/ACPBridge.swift. Lines 467 through 470 are the actual injection.

The agent process inherits ANTHROPIC_BASE_URL. The Anthropic SDK inside the subprocess uses whatever URL it sees in that variable. If the URL points at a local proxy that translates Anthropic's message shape to a Llama-serving backend, the rest of the agent loop is none the wiser. The tool-use frames go out, the tool results come back, the agent applies them to the user's screen via the macOS accessibility tree the way it always does. The Llama choice is downstream of one env var.

The Llama integration is not theoretical. Fazm has shipped explicit error-message handling for it. When a user's local server (LM Studio, Ollama, llama.cpp behind a proxy) reports an upstream error, the agent surfaces an actionable hint instead of a raw 500. Lines 2045 to 2058 of the same file detect the "No models loaded" case from LM Studio and the "connection refused" case from a stopped Ollama process.

Two lines of comment, one substring match, one actionable error. The Llama path is a first-class user surface, not a hidden setting. Which is also why the 2026 Meta pivot matters here: the people who installed this product specifically to keep their inference local are the people most exposed to a frontier-line freeze. The Llama 4 weights they already have are still good; the question of what to upgrade to next is suddenly less obvious.

The end-to-end path from a user typing a chat turn to a Llama running on the same MacBook

Spelled out as a sequence, the trip from a chat input to a local Llama 3 70B response takes five actors and roughly ten frames. The agent never knows the response came from Llama. The Llama proxy never knows it is running behind a desktop computer-use agent. The seam stays at one env var, and everything else is the standard Anthropic message shape.

The same diagram applies whether the "Local LLM" column is Llama 3 8B, Llama 3 70B, Llama 4 Scout, or some non-Llama open-weights model. Nothing about the upstream agent code changes when the underlying weights change. Which is also the answer to the underlying question this page's query implies, "what happens to my workflow if Llama 5 never ships." The workflow does not break. The integration was always at the protocol layer, not the model layer. What you lose is the upside that a bigger open-weights Llama would have brought, not the existing capability.

What is true, what is not true, and what is still pending

A clean status block for the four model labels people actually type into search bars, plus the rumour names that turn up in mid-2025 reporting and never resolve. Items with a check are confirmed against Meta's own posts. Items without are either unreleased or unconfirmed.

The two negatives at the bottom of that list are the part you cannot get from a content-mill page. Llama 4.X / 4.5 / Mango / Avocado are names that show up in second-hand reporting and have no Meta product page behind them as of this date. That does not mean they are vapour; it means they are not the answer to the literal query. The literal query has exactly one positive 2026 answer, and that is Muse Spark.

The honest counterargument: why this might still be wrong by the end of 2026

A few caveats that should make a reader less confident in the framing above. Meta has not formally discontinued Llama. The community license is still active, the back-catalogue is still on Hugging Face, and existing fine-tunes and quantisations are not going anywhere. Nothing in the Muse Spark launch post says "no more Llama frontier ever." A face-saving 4.X or 5 release later in 2026, perhaps with a smaller compute envelope and a tighter scope, is not ruled out. Meta has been here before with Llama 3.x point releases that arrived without a fanfare.

What is ruled out, by the post itself, is the "open-weights frontier from Meta" expectation that defined 2023 to early 2025. Muse Spark is closed. If the open-weights story matters to your stack, Meta is not the place to plan around anymore; the relevant horizon is whatever Mistral, DeepSeek, Qwen, and the other open-weights labs ship next, plus whatever 4.X-style update Meta might or might not bring out of the back catalogue. The Llama 5 specific question is still "no" on the public record either way.

For a Mac computer-use agent, the planning move that survives both outcomes is the one already in Fazm's code: keep the integration seam at one env var, keep the agent loop model-agnostic, and treat the model picker as something that follows the SDK's available-models payload at runtime. If Llama 5 ships next week, the same custom-endpoint field absorbs it. If it never ships, nothing changes downstream. That is the engineering value of having no opinion on the roadmap question while still being correct about today's status.