Llama 5 Is Meta's Best Release in Years, and Scout Is the One That Matters
Meta released three Llama 5 models on June 17: Scout, Maverick and Behemoth. The press cycle went almost entirely to Behemoth, because it is the largest open-weight model anyone has shipped and it scores 90.8 on MMLU. That is a fine headline. It is also, for most people reading this, completely irrelevant.
Behemoth is not something you run. It is something you rent from a partner, and once you are renting inference from a partner you are back in the same buy-versus-build conversation you were having about closed models, just with worse tooling. The interesting release is Scout.
What Scout Actually Does
Llama 5 Scout carries a 1M token context window, accepts images and video frames, and runs at usable speed on a single 24 GB consumer GPU. Not quantised into uselessness. Not with a 4K context. A full million-token window on hardware that costs less than a mid-range laptop did five years ago.
Benchmark-wise it is not going to threaten anything at the top of our LLM leaderboard. It posts 80.2 on MMLU and 39.8 on SWE-bench Verified, which puts it roughly where hosted frontier models were in early 2025. The point is not the score. The point is where the score is running.
The Three Models, Briefly
- Scout: 1M context, vision, single-GPU. The local and edge play.
- Maverick: 2M context, mixture-of-experts, the general purpose middle. Hosted at roughly $0.35 per million input tokens.
- Behemoth: 2M context, largest open weights released to date, partner-hosted only. Frontier-adjacent reasoning at $1.20 per million input.
All three ship under a revised Llama 5 Community License. The revision matters: the monthly-active-user threshold that triggered the commercial licence requirement moved up substantially, which puts almost every startup safely outside it.
Why Local Suddenly Got Interesting Again
The argument against local inference has always been the same: whatever you can run at home is so far behind the hosted frontier that the cost saving is not worth the quality loss. That argument holds right up until the point where local quality crosses the threshold for a specific job.
Scout crosses it for a lot of jobs. Document classification. Extraction from scanned forms. First-pass code review. Log triage. Anything where you are running the same narrow task ten thousand times a day and the marginal quality difference between Scout and a frontier model rounds to noise, but the cost difference does not.
Run the numbers yourself with our RAM calculator to see what quantisation level fits your card, then compare the hosted equivalent on the price comparison tool. For high-volume narrow tasks the crossover point is usually somewhere around two million requests a month.
What Meta Is Actually Doing Here
The strategic read is unchanged from Llama 4. Meta does not need to sell inference. It needs the ecosystem to standardise on weights it controls, so that every tool, every fine-tuning pipeline and every deployment guide assumes Llama by default. Giving away a genuinely good single-GPU model is the cheapest way to buy that.
It is working. Scout was in the default model list of half a dozen local runners within a week of release, which is faster adoption than any previous Llama point release.
The Catch
Vision on Scout is real but shallow. It handles document layout and simple charts well and falls apart on anything requiring fine spatial reasoning. Video frame input is technically supported and practically unusable at more than a few frames per request. And the 1M window, as with every 1M window, degrades in the middle. Treat it as a 300K window with a long tail you can reach into occasionally, not as a place to dump a monorepo.
Compare Scout and Maverick side by side on our model comparison tool, or browse the full model list for hosted pricing across providers.