TL;DR: Muse Spark, launched April 8, 2026, is the first Meta flagship you cannot download — API and Meta AI app only, no weights, no self-hosting path. Llama 4 remains downloadable under its Community License, but Meta’s hosted Llama API is dead and no Llama 5 is announced. The practical move is diversifying to Apache 2.0 and MIT models that already outperform Llama 4.

	Keep running Llama 4	Migrate to Qwen / Gemma / DeepSeek	Use Muse Spark via API
License type	Llama 4 Community License (700M MAU cap, attribution, EU multimodal restriction)	Apache 2.0 / MIT — no caps, no attribution	Proprietary, usage-based ToS
Data sovereignty	Full — weights on your hardware	Full — weights on your hardware	None — inference on Meta servers
Cost	$0 after hardware	$0 after hardware	Paid API (opened July 9, 2026)
The catch	Stagnant family: no 4.1/4.5/5 roadmap, hosted API shut down	Migration effort: re-test prompts, re-pull models	No privacy, no offline, terms can change

Honest take: If your stack has any Meta-only dependency, replace it this quarter — not because Llama 4 stopped working, but because the models that replaced it are better and carry cleaner licenses. Qwen3.6-35B-A3B on a 24GB card is the single best migration target.

Meta spent four years telling everyone that open weights were the point. Llama 1 leaked, Llama 2 and 3 shipped to Hugging Face within days of announcement, and Mark Zuckerberg published “Open Source AI is the Path Forward” in 2024 as a manifesto. Then, on April 8, 2026, Meta Superintelligence Labs shipped Muse Spark — a multimodal reasoning model with no downloadable weights, no Hugging Face repo, and no self-hosting story at all.

If you run open-source AI on your own hardware, this is worth ten minutes of your attention. Not because your Ollama setup breaks — it doesn’t — but because the risk profile of building on Meta’s ecosystem just changed, and the alternatives happen to be in the best shape they’ve ever been.

What actually happened, April through July

The timeline matters, because Muse Spark wasn’t a one-off experiment. It’s a pattern now:

• April 8, 2026: Muse Spark launches as the first model from Meta Superintelligence Labs (MSL), the division built under Alexandr Wang after Meta shelved Behemoth, the planned 2-trillion-parameter Llama 4 teacher. Access: the Meta AI app and meta.ai. Weights: none. API: invitation only.
• July 6, 2026: Meta’s hosted Llama API shuts down. The official cloud path to Llama models is gone.
• July 7, 2026: Muse Image launches and Muse Video is previewed — MSL’s first in-house media generation models. Both proprietary, shipped as product features inside Meta AI, Instagram, and WhatsApp. No weights.
• July 9, 2026: Muse Spark 1.1 opens Meta’s first paid, OpenAI-compatible API.

Three releases, zero downloads. Meta has said it hopes to release future versions under an open license, framing the closure as temporary. Maybe. But “we’ll open it later” is a hope, not a license file, and self-hosters plan around license files.

What “closed” means in practice

For anyone used to the Llama routine — announcement, then weights on Hugging Face within days to weeks — here is the concrete difference with Muse Spark:

• No ollama pull, ever. There is no GGUF, no quantization community, no Unsloth dynamic quants three days after release. The model exists only behind Meta’s endpoints.
• No privacy path. Every prompt goes to Meta’s servers. For the workloads that drive most self-hosting — internal documents, proprietary code, health data, anything under GDPR or client confidentiality — Muse Spark is simply not an option, at any price.
• No version pinning. Meta can update, throttle, or retire the model under you. Self-hosters who lived through cloud model deprecations know exactly what this feels like.
• Free tier ≠ open. Muse Spark is free to use inside the Meta AI app. That’s a distribution strategy, not openness. You are the training-feedback loop.

One distinction the coverage keeps blurring, and this site won’t: Muse Spark is a new model family, not a Llama replacement. Llama 4 Scout (109B total, 17B active, 16 experts) and Maverick (400B total, 17B active, 128 experts) remain downloadable on Hugging Face as of mid-August 2026. Nothing was retracted. The Llama 4 Community License still applies — and note it is not Apache 2.0, despite what several posts claim: it carries a 700M-monthly-active-user cap, a “Built with Llama” attribution requirement, and an EU restriction on multimodal use. Our open-source LLM license shootout covers the fine print.

Is Llama 4 still safe to run?

Yes — and also, that’s the wrong question. The weights on your disk can’t be un-released. If Llama 4 Scout serves your workload today, it serves it tomorrow regardless of anything Meta does.

The right question is whether to keep investing in the Llama line — fine-tunes, prompt libraries, eval suites, infrastructure assumptions. And there the signals are hard to ignore:

• The hosted Llama API is gone as of July 6, 2026.
• No Llama 4.1, 4.5, or 5 exists on any first-party source, sixteen months after Llama 4 shipped.
• Behemoth was cancelled and its team redirected into a proprietary lab.
• Eleven of the fourteen researchers on the original 2023 Llama paper have left Meta.

Does the Muse Spark precedent put a hypothetical Llama 5 at risk of going closed? Honest answer: unknown. Meta says it has no plans to close Llama. But Muse Spark proves the policy changes when commercial pressure demands it, and the org that made Llama open no longer exists in recognizable form. Treat future Llama releases as a pleasant surprise, not a roadmap.

The audited alternatives, by workload

Here’s the part that makes this an easy article to write: in August 2026, dropping Meta costs you nothing. Every entry below has had its license verified against the actual repository or model card for our previous coverage — not against a press release.

General chat and reasoning

Qwen3.6-35B-A3B (Apache 2.0) is the practical champion for single-GPU self-hosting: a 35B MoE with 3B active parameters, roughly 21GB at Q4_K_M, which fits a 24GB card like an RTX 3090 with room for context. It posts 73.4% on SWE-bench Verified — well clear of Llama 4 Scout. Setup notes in our Qwen3.6-35B-A3B local guide.

Gemma 4 (Apache 2.0 — the first Gemma generation under a real OSI license, released March 2026) covers the smaller tiers: the 12B QAT build runs in about 7GB, and the 26B-A4B MoE QAT fits in roughly 15GB on a 16GB card like an RTX 4080. See the Gemma 4 QAT self-hosting guide for the quantization trap to avoid.

Coding

Codestral 2 (Apache 2.0, relicensed from Mistral’s non-production license in April 2026) is a 22B dense FIM specialist that runs in ~13GB — the clean pick for a 16GB card. DeepSeek V4-Flash (plain MIT, 284B/13B-active) is the heavyweight: self-hostable from ~103GB at 3-bit if you have the RAM, or dirt cheap via API. GLM-5.2 (MIT, 744B MoE) tops SWE-bench Pro at 62.1% but is datacenter-scale; our GLM-5.2 review does the real VRAM math. If you want frontier-class coding without owning the hardware, renting a pod on RunPod beats any Meta option on both privacy terms and license clarity. Watch Kimi K2.7 licensing, though — it’s Modified MIT with a commercial clause above 100M MAU or $20M monthly revenue.

Image generation

Meta licensing Midjourney and Black Forest Labs for years, then shipping proprietary Muse Image, changes nothing locally: FLUX.1 Schnell is Apache 2.0 and runs on consumer VRAM. Be precise here — FLUX.1 Dev is non-commercial, a distinction the queue-jumping listicles routinely get wrong. SDXL’s LoRA ecosystem remains the widest, and SD 3.5 Medium runs on 8GB cards.

Speech

Kyutai Pocket TTS (MIT, code and weights) clones a voice from a single WAV file on a laptop CPU — reviewed here.

The license table

Model	Org	License	Commercial use	Consumer GPU?
Qwen3.6-35B-A3B	Alibaba	Apache 2.0	✅ Unrestricted	✅ 24GB
Gemma 4 (12B / 26B-A4B)	Google	Apache 2.0	✅ Unrestricted	✅ 8–16GB
Codestral 2	Mistral	Apache 2.0	✅ Unrestricted	✅ 16GB
DeepSeek V4-Flash	DeepSeek	MIT	✅ Unrestricted	⚠️ 128GB+ RAM offload
GLM-5.2	Z.ai	MIT	✅ Unrestricted	❌ Datacenter
FLUX.1 Schnell	Black Forest Labs	Apache 2.0	✅ Unrestricted	✅ 12GB
Llama 4 Scout / Maverick	Meta	Llama 4 Community License	⚠️ 700M MAU cap + attribution + EU limits	✅ / ⚠️
Muse Spark / Image / Video	Meta	Proprietary	❌ API terms only	❌ No weights

What this moment is actually telling you

The tempting narrative is “Meta killed open-source AI.” The evidence says the opposite. When Llama launched, Meta was the open-weight ecosystem. In August 2026, Chinese and European labs ship MIT and Apache 2.0 models that beat Llama 4 on coding and reasoning benchmarks, Google put Gemma under Apache, and the tooling — Ollama, vLLM, llama.cpp — is model-agnostic by design. Meta walking away from open weights in 2026 is like a founder leaving a company that already IPO’d. We made the longer version of this argument in the sovereignty case for open-source AI.

The actionable lesson is narrower: never let one vendor be your stack. That means model-agnostic serving (Ollama or vLLM, not a vendor SDK), prompts and evals you can re-run against a new model in an afternoon, and a license check before a model becomes a dependency — the check takes five minutes and this site’s reviews always lead with it.

For hardware sizing on the alternatives above, our sister site has the GPU guides at runaihome.com, and if you’re pointing a coding agent at these models, aicoderscope.com covers wiring them into Cline and Cursor as BYOK backends.

FAQ

Can I still download Llama 4? Yes. Scout and Maverick remain on Hugging Face under the Llama 4 Community License, and nothing suggests removal. The license allows commercial use below 700M monthly active users with attribution; EU users face restrictions on the multimodal variants. Existing downloads are yours regardless of future policy.

Is Muse Spark free? Free to use inside the Meta AI app and on meta.ai. The OpenAI-compatible API that opened July 9, 2026 is paid. Neither tier includes weights, self-hosting, or any offline path — “free” here means Meta’s servers process your prompts.

What’s the closest self-hostable equivalent to Muse Spark? For multimodal reasoning on your own hardware, Gemma 4 (text, image, and audio input, Apache 2.0) is the closest single-model answer at consumer scale, and Qwen3-VL covers vision workloads across sizes from 2B up. Nothing local matches a frontier API model’s ceiling — that gap is the price of sovereignty, and it narrowed again this year.

Sources

• The Next Web — Meta’s Muse Spark is here – and it’s closed source
• Meta AI Blog — Introducing Muse Image and Muse Video
• CNBC — Meta debuts Muse Image, Superintelligence Labs’ first AI image model
• gHacks — Meta Launches Muse Spark, Its First Proprietary AI Model
• AI News — Did Meta Sacrifice Its Open-Source Identity?

Recommended Gear

• RTX 3090 — 24GB VRAM, the used-market sweet spot for Qwen3.6-35B-A3B at Q4_K_M
• RTX 4080 — 16GB VRAM, comfortable for Gemma 4 26B-A4B QAT and Codestral 2

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe