Llama 3.1 70B Instruct in 2026: When an Old Model Still Fits

Llama 3.1 70B Instruct is more than two years old, and on raw capability it has been overtaken several times, including by Meta's own Llama 3.3 70B at the same size. It is still listed by most routers and still runs in a lot of production systems. Neither fact is an accident, and neither is a reason to choose it for new work without thinking.

This guide covers what the model is, where its successor beat it, the honest reasons it is still served, and what its licence requires of you, which is more than many teams realise.

What it is, in 2024 terms

Meta released the Llama 3.1 family on 23 July 2024: an 8B, a 70B and the 405B flagship. Meta's model card describes the 70B as a dense transformer using grouped-query attention, pretrained on about 15 trillion tokens of public data, with a knowledge cutoff of December 2023 and a 128K context window. It officially supports eight languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.

Meta reports 7.0 million H100 GPU hours to train the 70B. The instruct model supports tool use through its chat template, and Meta's launch post described the upgraded 8B and 70B as having stronger tool use and reasoning than the Llama 3 models before them.

The model card's headline numbers for the instruct version: MMLU 83.6, MMLU-Pro (CoT) 66.4, HumanEval 80.5, MATH (CoT) 68.0, GPQA 46.7, IFEval 87.5 and BFCL 84.8. Those were strong for an open model in mid-2024. Against 2026 open-weight models they are modest, and there is no reasoning mode or effort control at all.

Llama 3.3 70B beat it at the same size

Meta released Llama 3.3 70B Instruct on 6 December 2024, same parameter count, same 128K window. Its model card compares the two directly:

• MMLU-Pro (CoT) — 68.9 against 66.4
• HumanEval — 88.4 against 80.5
• MATH (CoT) — 77.0 against 68.0
• GPQA Diamond (CoT) — 50.5 against 48.0
• IFEval — 92.1 against 87.5
• BFCL v2 — 77.3 against 77.5

Tool calling is the one place 3.1 holds level. Everywhere else 3.3 is better at identical hardware cost. If you are choosing a 70B Llama for a new project with no other constraint, the decision is already made.

Notice also that the 3.1 card reports BFCL 84.8 and GPQA 46.7, while the 3.3 card reports 3.1 at BFCL v2 77.5 and GPQA Diamond 48.0. Same model, different benchmark versions and settings. It is a small, clean example of why a number lifted from one table should never be compared with one from another.

The reasons it is still served

Three of them hold up:

• Fine-tunes built on it. A large ecosystem of derivatives starts from 3.1 70B, including Hermes 3 70B and several roleplay and creative-writing tunes. If you depend on one, you depend on its base.
• Pinned evaluations. Systems validated against 3.1 in 2024 have test suites, prompts and output parsers tuned to its behaviour. Switching models is a project, not a config change.
• Predictable, cheap hosting. It is one of the most widely served open models, so it is available from many providers. OpenRouter lists it at $0.40 per million tokens for both input and output at the time of writing, with up to 16,384 output tokens.

Flat input and output pricing is worth noticing. Many current models charge several times more for output than input, so for generation-heavy work such as drafting or summarising into long form, a flat rate narrows the gap more than the headline input price suggests.

The licence is not MIT

Llama 3.1 ships under the Llama 3.1 Community License, a custom licence. It permits commercial use, but with conditions that bite at scale or on redistribution:

• If your products exceed 700 million monthly active users, you must request a separate licence from Meta, which Meta may grant at its discretion.
• If you distribute the model or a product built on it, you must prominently display "Built with Llama".
• If you use its outputs to create or improve an AI model that you distribute, that model's name must begin with "Llama".
• Use is bound by Meta's Acceptable Use Policy.

Meta's launch post did loosen one thing: outputs from Llama 3.1 models, including the 405B, may be used to improve other models, which Meta presented as a change from earlier terms. The naming condition still applies when you distribute the result. Open-weight model licences compared puts these terms next to MIT and Apache alternatives.

Running it yourself

Seventy billion dense parameters means every token touches all of them. In BF16 the weights alone are roughly 140GB, so you need multiple 80GB cards or a quantised build. A 4-bit quantisation brings the weights to around 35 to 40GB, which fits on a single 48GB card with a short context, or an 80GB card with room for a longer KV cache.

That is the main thing 3.1 70B still offers over hosted frontier models: a known, stable, widely tooled model you can run on hardware you control. It is also exactly what 3.3 70B offers at the same footprint, so the self-hosting case for 3.1 specifically comes back to fine-tunes and pinned behaviour.

A decision rule

1. New project, need a 70B Llama: use 3.3 70B.
2. Depend on a fine-tune built on 3.1 70B: stay, and watch whether the tuner publishes a 3.3 or newer base.
3. Existing system validated on 3.1: keep it pinned, build a regression set, and schedule a migration rather than letting a provider deprecation schedule it for you.
4. Need reasoning, long-horizon agents or a context beyond 128K: neither 70B Llama is the right tool in 2026.