Mistral Large 4 vs Llama 4 Maverick: A Generation Gap, Not a Price Gap

Mistral Large 4 and Llama 4 Maverick are the same kind of object and not the same generation of object. Both are sparse mixtures of experts with a one-million-token context window on the model card, both take images as well as text, and both can be downloaded and run on your own hardware. Maverick shipped on 5 April 2025 with 400 billion total parameters and 17 billion active. Mistral Large 4 previewed on 6 October 2026 and published its weights on 29 October, with 1.05 trillion total and 52 billion active. Nineteen months separate them. At the median price across Maverick's hosting market, only about 1.3 times the cost separates them too — which is not what almost anyone expects to hear about a model that old.

The short answer, before the caveats

If you are starting a new build in late 2026 and you need reasoning, long-document work or agentic coding, Mistral Large 4 is the better model and the more expensive one. If you need an open-weight model you can fine-tune, run air-gapped, or call for a fraction of a cent per request on a simple extraction task, Llama 4 Maverick is still a defensible choice — not because it is competitive on capability, but because the workload may not need capability. The mistake is treating a 19-month gap as a rivalry. It is a price tier that happens to share a category, and the price part is stranger than the capability part.

The numbers side by side

• Total parameters — Mistral Large 4 is 1.05T across 128 routed experts plus one shared expert, vs 402B on Llama 4 Maverick across 128 experts, listed as 400B by Meta.

• Active parameters — 52B per token on Mistral Large 4 (49B for the language stack, counting embeddings and output layers brings it to 52B) vs 17B on Llama 4 Maverick, so the per-token compute gap is about three to one.

• Context — 1M tokens advertised on both; what a host actually serves is a different number, and a later section is about that.

• Modality — text and image in, text out, on both. Mistral Large 4's vision encoder is 1.6B parameters; Maverick's early-fusion encoder is not broken out in Meta's published figures.

• Reasoning — Mistral Large 4 exposes a low, medium and high effort ladder with high as the default. Artificial Analysis's specification panel lists Llama 4 Maverick as having no reasoning mode at all.

• Price per million — $0.70 input and $2.10 output from Mistral directly, with cached input at $0.07, vs no first-party list price for Maverick at all, because Meta does not sell it. Hosts charge as little as $0.17 input and $0.60 output; Artificial Analysis puts the provider median at $0.50 and $1.68.

• Licence — the Mistral Ecosystem Open-Weight License, a modified MIT with a model-as-a-service carve-out, vs the Llama 4 Community License with its 700-million-monthly-active-user threshold.

• Release — 29 October 2026 for the Mistral Large 4 checkpoint, against 5 April 2025 for Llama 4 Maverick, with a knowledge cutoff of August 2024.

The benchmark fight is between two different rulers

Search this matchup and you will find pages claiming Mistral Large 4 scores 38.0 on the Artificial Analysis Intelligence Index against Maverick's 14.3, and other pages claiming Maverick wins nine of eleven benchmarks. Both sets of numbers are real in the sense that somebody printed them. Neither is a comparison.

On the current Artificial Analysis Intelligence Index v5, rebuilt on 27 October 2026, Mistral Large 4 scores 26 at its high reasoning setting and 23 at medium. The 38.0 you see quoted is its score on the previous v4.3.2 build, from the October preview, and it is not subtractable from anything measured after the rebuild — the index changed, not just the model. Maverick, meanwhile, has no v5 index score at all. Its Artificial Analysis page shows "N/A" where the intelligence figure should be, alongside a speed of 62.7 output tokens per second and a note that it is slower than average for its class. The 14.3 figure floating around comparison sites predates the rebuild.

What is measurable is narrower and more honest. Mistral reports 41% on Terminal-Bench 4.0, up from 28% at preview; 61.7% on DeepSWE v1.1; 79% on CyberGym-E2E; and 51 on the Artificial Analysis Cyber Index, third among open-weight models. Every one of those except the Cyber Index is vendor-reported on Mistral's own agent scaffold, and the Terminal-Bench figure in particular has been reproduced much lower — 24.7% — when the same weights run on Artificial Analysis's neutral mini-swe-agent harness. Maverick's published numbers are from April 2025 and use an older suite entirely: 85.5 MMLU, 80.5 MMLU-Pro, 69.8 GPQA Diamond, 91.6 DocVQA, all Meta-reported at launch. There is no shared harness in which the two have been run, so the only defensible statement is directional: the newer, larger, reasoning-trained model is better at hard tasks, by an amount nobody has cleanly measured across the pair.

What a million tokens actually buys you on each

The context window is where the two models look most alike and behave most differently in production.

Llama 4 Maverick's model card says 1,048,576 tokens, and Artificial Analysis's spec panel repeats 1M. In practice, most hosts cap it far lower — 128K is common — and almost none will give you both the full window and a useful output budget. Maximum output on Maverick is typically 8K to 16K depending on the provider, and Meta's own hosted endpoint has been listed at roughly 4K. So a "1M context" build on Maverick is usually a 128K build on a third-party host, and you find that out at the first long document.

Mistral Large 4 publishes 1M through Mistral's own API, which is at least a single accountable number rather than a per-host lottery. Mistral's documentation does not publish a maximum output figure, so the same advice applies: measure the pair — input window and output cap — on the endpoint you will actually call, not on the card.

This is the sort of thing a router is for. OrcaRouter passes provider list price through at 0% markup and fails over automatically between providers, so a host that quietly trims a context window shows up as a routing decision rather than a production incident. Neither Mistral Large 4 nor Llama 4 Maverick is in our catalogue today, so that is an argument about how to run whichever open-weight model you pick, not a claim that we serve these two.

The bill, worked out — and why the spread is the story

Take a modest production workload: 100 million input tokens and 40 million output tokens a month.

• Mistral Large 4 at list — 100 × $0.70 is $70, and 40 × $2.10 is $84, for $154 a month.

• Llama 4 Maverick at the provider median — 100 × $0.50 is $50, and 40 × $1.68 is $67.20, for $117.20 a month.

• Llama 4 Maverick at a competitive host — around $0.25 and $0.92 puts the same workload at $61.80.

• Llama 4 Maverick at the cheapest hosts — around $0.17 and $0.60 puts it at $41.

So the ratio against Mistral Large 4 runs from 1.3 times at the median to about 3.8 times at the bottom of the market. That range is not a property of the model. It is a property of where you buy Maverick, and it is wider than the gap between the two models at the middle of the distribution. A team that benchmarks on the cheap host and budgets on the median has built a 3× error into its own forecast.

Set that against capability you cannot currently measure pair-wise, and the decision stops being arithmetic. The reasonable framing is cost per completed job, not cost per token: if Mistral Large 4 finishes in one pass a task that needs three attempts and a repair loop on Maverick, even the 1.3× median gap disappears. That is a claim you have to test on your own tasks, and it is the single most valuable experiment either model can be put through this month.

The thing nobody says out loud: Maverick is the last open Llama

Llama 4 Maverick has not been deprecated. It is still on Hugging Face, still under the Llama 4 Community License, and still widely deployed for on-premises and air-gapped work. But it is the end of a line. Meta's frontier development moved to the closed Muse Spark family in 2026, and no Llama 5 has shipped — the pages claiming one did carry specifications that contradict each other and Meta's own announcements. Investment in the open Llama track has effectively stopped.

That changes what "pick Maverick" means. It is not picking a model that will improve; it is picking a model that will stay exactly as it is while the tooling around it slowly drifts. For a fine-tuned, frozen, deployed system that is fine — arguably an advantage, because nothing moves under you. For a platform you intend to grow on, it is a dead end with a two-year runway. Note that the knowledge cutoff is August 2024: the model has never seen anything that happened after that, and no amount of prompt engineering will fix it.

Mistral Large 4 has the opposite risk profile. It is six days old as an open-weight release, its documentation still labels it a public preview, and its licence requires a commercial agreement if you intend to resell inference on it. Young and moving, with terms that need reading before you build a business on top.

Where Maverick still wins

Three cases, and they are not small ones.

The first is hardware. Maverick's 400B parameters fit in roughly 373 GB at FP8, which is a single eight-GPU H100 or H200 node. Mistral Large 4's FP8 checkpoint is about 1.06 TB across 54 shards, and Mistral says the quantised NVFP4 build runs on four B200 or B300 GPUs. If your deployment is one node in a rack you already own, the choice has been made for you.

The second is fine-tuning. Maverick has had nineteen months of community LoRAs, quantisation recipes and serving guides. Mistral Large 4 has had six days. The ecosystem gap is real and it is measured in debugging hours, not benchmarks.

The third is task shape. Classification, extraction, routing, summarisation of short documents, image captioning at volume — none of these need a trillion-parameter reasoning model, and all of them get cheaper on a 17B-active one. Paying a premium for reasoning tokens you never use is the most common way teams waste money on model selection.

If you want a current open-weight model that is neither of these, the OrcaRouter catalogue carries Kimi K3 and GLM-5.3-Flash, both callable from the same API key as the rest of the 200-plus models we route, at provider list price with no markup. They bracket this comparison rather than sitting inside it — GLM-5.3-Flash lists at $0.15 input and $0.50 output, well below either model here, while Kimi K3 lists at $3.00 and $15.00, well above both. Neither is a substitute for the two in this piece, but they are the two obvious places to look if neither end of it fits.

How to decide in an afternoon

Pull twenty real tasks from your logs. Run them through both models with the same prompts and the same output cap, and score them on whether the output was usable without a human edit. Record the tokens each one burned, multiply by the prices above, and divide by the number of usable outputs. That gives you cost per usable result, which is the only number that survives contact with a budget.

My expectation, stated as an expectation rather than a finding: Mistral Large 4 wins the hard half of that set and loses the easy half on cost, and the split point lands wherever your tasks stop needing multi-step reasoning. Maverick is not the better model. It is the cheaper one at the median for as long as cheap is enough — and after nineteen months with no successor, that is a bet with a visible expiry date.