benchr
Lead with the documented ceiling: Meta lists a 10-million-token context for Llama 4 Scout. Scout has 109B total parameters, so an ideal 4-bit weight calculation is about 54.5GB before quantization metadata and runtime buffers. That can make an 80GB accelerator a candidate for weight fit, but it does not establish that a full 10M-token request fits: KV cache, cache precision, concurrency, and the serving implementation add separate memory requirements. Treat Scout as a long-context candidate to size and reproduce, not a one-GPU guarantee.
Here's the turn. Llama 4 shipped on April 5, 2025. In AI-model years that's old, and Meta has quietly stopped treating it as the future. On April 8, 2026, Meta Superintelligence Labs announced a new closed-source model, Muse Spark, and in that announcement it refers to "Llama 4 Maverick" as "our previous model." Read that plainly: the company that built the most-downloaded open-weight family has moved its flagship behind a closed door. Llama 4 isn't being iterated at the frontier. It's the last major open-weight Llama, and you should buy into it knowing that.
This review covers the herd as it exists today, told through the three variants Meta named, and it's honest about what each one is for now that the generation is aging. Every number below is from Meta's official pages, and the benchmark figures are Meta's own: vendor-reported, not independently reproduced.
The herd, variant by variant
Llama 4 was Meta's first natively multimodal, Mixture-of-Experts (MoE) generation. Meta announced three sizes. Two shipped as open weights; one didn't ship at all. Knowing which is which is the whole buying decision, so start here.
| Variant | Status | Context | Parameters | Best use |
|---|---|---|---|---|
| Llama 4 Scout | Released, open weights | 10M tokens | 17B active / 109B total, 16 experts | Very-long-context evaluation with verified memory fit |
| Llama 4 Maverick | Released, open weights | 1M tokens | 17B active / 400B total, 128 experts | Top open-tier reasoning and multimodal |
| Llama 4 Behemoth | Announced, never released | Not specified | 288B active / ~2T total, 16 experts | N/A; previewed as "still training" |
Scout is the long-context variant. Its 17B active parameters reduce compute per token, while all 109B parameters still matter for weight storage. A suitable Int4 checkpoint may fit its weights on an 80GB accelerator; the usable context then depends on KV-cache memory, runtime overhead, concurrency, and serving support. If your workload uses enormous documents, codebases, or transcripts, reproduce the target length and retrieval quality on the exact deployment before treating the 10M ceiling as usable capacity.
Maverick is the heavier sibling: same 17B active parameters but 400B total across 128 experts, with a 1M-token context per the official Hugging Face release blog. It's the one Meta positioned as its strongest open chat-and-reasoning model, and it's the one Meta now calls its "previous model." Both Scout and Maverick are downloadable from llama.com and Hugging Face, and both were wired into Meta AI across WhatsApp, Messenger, Instagram Direct, and meta.ai.
Behemoth is the asterisk. At a reported 288B active parameters and nearly two trillion total, it was the teacher model the others were said to distill from — and it was only ever previewed as "still training." Meta never released it. No official cancellation has been published either; third-party outlets have reported a pause for performance reasons, but that isn't confirmed by Meta. Treat Behemoth as a name on a slide, not a model you can use.
The license clause that decides who can use it
Llama 4 is "open weight," not open-source, and the distinction is in the license. The Llama 4 Community License lets almost anyone download, run, fine-tune, and ship commercially, with attribution and a "Built with Llama" requirement plus an acceptable-use policy. For the vast majority of teams that's clean enough to deploy with confidence. There's one line that matters if you're large, though.
For everyone below that line, the license is permissive enough that the comparison with truly open competitors comes down to capability and support, not legal risk. benchr's survey of the open-weight tier right now places Llama 4 against the Apache- and MIT-licensed alternatives, and also that some of those, like the models in the Mistral review, carry their own research-only or restricted terms. "Open" is a spectrum, and Llama 4 sits in the permissive-but-conditional middle of it.
The benchmarks, read as vendor figures
Meta published a full benchmark sheet at launch, and it's the source for every score people quote. Read all of these as Meta's own reported numbers: vendor-reported, not independently reproduced. On MMLU Pro, Meta lists Maverick at 80.5 and Scout at 74.3. On the multimodal MMMU, Maverick is 73.4 and Scout 69.4. Document and chart understanding looked strong, with both variants at 94.4 on DocVQA and Maverick at 90.0 on ChartQA. Coding was the soft spot Meta's own sheet shows: LiveCodeBench came in at 43.4 for Maverick and 32.8 for Scout, well behind dedicated coding models.
One figure needs a warning label. Meta's blog cited an LMArena Elo of 1417 for Maverick, but that was an experimental, chat-optimized variant, not the released weights, and it drew criticism precisely because the version that topped the leaderboard wasn't the one you could download. Don't carry the 1417 into a comparison with shipped models. For why a single headline score is a weak basis for any decision, benchr's piece on why the benchmarks stopped telling you anything is the relevant background.
What it costs to run
There's no Meta price to quote, because Meta doesn't sell first-party API access for Llama 4. It ships the weights and you run them, or you call a third-party host that prices separately. Meta's own blog does cite an estimated inference cost for Maverick of roughly $0.19 per million tokens (blended 3:1 input-to-output, with distributed inference). Read that exactly as written: it's a cost estimate Meta published, not a price Meta charges you. Your real bill is hardware and operations, or whatever a host like IBM watsonx.ai or a cloud provider lists.
That makes the practical question the same one every open-weight model raises: self-host or rent. Scout is smaller than 400B-parameter Maverick, but model-weight fit and full-context fit are different calculations. Size the 109B weights, KV cache, runtime buffers, concurrency, and target context separately; a single high-end GPU may hold a quantized checkpoint without supporting a full 10M-token request. A hosted endpoint also may expose a lower cap, so verify the deployed limit rather than the model-card ceiling.
The verdict
Llama 4's 4.0 is an editorial score, not a benchmark. It reflects Scout's advertised 10M-token context and downloadable weights, while marking down the generation's age and the lifecycle uncertainty created by Meta describing Maverick as its “previous model.” Check the current license, memory fit, support path, and retrieval accuracy on your own deployment.
Go with Llama 4 Scout when a giant context on hardware you control is the requirement and you can absorb that the generation won't get better. Pick Maverick when you want the strongest open-tier reasoning Meta shipped and 1M tokens is enough. Skip Behemoth entirely, because it doesn't exist to download. And if your roadmap needs frontier accuracy, a vendor actively iterating, or freedom from the 700M-MAU clause, look past Llama 4 — it was the high-water mark for open Llama, and the tide has gone out.
Frequently asked
Is Llama 4 free to use?
The weights are free to download. Llama 4 Scout and Llama 4 Maverick are published under the Llama 4 Community License on llama.com and Hugging Face, so you can self-host at no licensing fee. It is not OSI open-source, and there is a catch: if your products had more than 700 million monthly active users in the month before Llama 4's release, you must request a separate license from Meta, granted at its sole discretion. Meta does not sell first-party API access, so the running cost is your own hardware or a third-party host's per-token rate.
What is Llama 4 Scout's context window?
Scout has a Meta-documented 10,000,000-token context ceiling, while Maverick is listed at 1,000,000 tokens. The 10M number does not guarantee retrieval quality or that a full-length request fits on one GPU. An Int4 checkpoint may fit Scout's weights on an 80GB accelerator, but KV cache, runtime overhead, cache precision, and concurrency must be sized separately for the target context.
What happened to Llama 4 Behemoth?
Behemoth (288B active, nearly two trillion total parameters, 16 experts) was previewed in April 2025 as still training and was never officially released. As of late May 2026, Meta has not published an official cancellation. Third-party outlets have reported it was paused or held back for performance reasons, but treat that as unverified by Meta. For planning purposes, Behemoth is not a model you can download or call.
Is Llama 4 still worth building on in 2026?
It depends on what you value. Llama 4 is now an aging generation: on April 8, 2026, Meta announced a closed-source model, Muse Spark, that names Llama 4 Maverick as its previous model, signaling Meta has moved past open-weight Llama as its flagship. If you need a permissive, self-hostable model with a very large context and a license clean enough for almost any company, Llama 4 still does that job. If you need frontier accuracy or a vendor on a clear update path, it is being left behind.
Are Llama 4's benchmark scores reliable?
They are Meta's own reported figures. Numbers like Maverick's 80.5 on MMLU Pro and Scout's 74.3 come from llama.com and should be read as vendor-reported, not independently reproduced. One number deserves extra caution: the 1417 LMArena Elo Meta cited was an experimental chat-tuned variant, not the released weights, and it drew criticism for not matching what shipped. Benchmark Llama 4 on your own tasks before trusting any single headline score.
Changelog
- July 24, 2026 — Separated single-accelerator Int4 weight fit from the KV-cache and runtime memory needed for a full 10M-token request; removed one-GPU full-context claims.
- July 23, 2026 — Labeled the rating as editorial and replaced universal context, license, hardware, and lifecycle conclusions with checks to run before deployment.
- May 30, 2026 — Originally published. Variant lineup, context windows, parameter counts, license terms, and benchmark figures verified against Meta's official Llama 4 blog (ai.meta.com), llama.com model and license pages, and the meta-llama Hugging Face model cards and release blog. Benchmark scores are labeled Meta-reported; the 1417 LMArena Elo is flagged as an experimental chat variant. Successor status confirmed against Meta's April 8, 2026 Muse Spark announcement, which names Maverick as its "previous model." No official Behemoth cancellation was found.
References
- Meta, "The Llama 4 herd: the beginning of a new era of natively multimodal AI innovation," ai.meta.com/blog/llama-4-multimodal-intelligence, April 2025.
- Meta, "Llama 4 models," llama.com/models/llama-4, accessed May 2026.
- Meta, "Llama 4 Community License Agreement," llama.com/llama4/license, effective April 5, 2025.
- Meta, "Llama-4-Scout-17B-16E-Instruct model card," huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct, accessed May 2026.
- Meta, "Llama-4-Maverick-17B-128E-Instruct model card," huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct, accessed May 2026.
- Hugging Face, "Welcome Llama 4 Maverick & Scout on Hugging Face," huggingface.co/blog/llama4-release, April 2025.
- Meta, "Introducing Muse Spark," ai.meta.com/blog/introducing-muse-spark-msl, April 8, 2026.