Llama 3.3 70B Instruct: A Post-Training Upgrade Worth Knowing

Llama 3.3 70B is easy to misread as a new generation. It is not. It is the same size and the same broad architecture as Llama 3.1 70B, released about four and a half months later with better post-training. Meta shipped it on 6 December 2024 as a single text-only 70B model; there was no 3.3 at 8B or 405B.

That framing matters because it tells you what to expect. A post-training update moves instruction following, code and maths. It does not usually move raw knowledge, and the benchmark table shows exactly that split.

What stayed the same

The basics carry over from Llama 3.1. Pretraining used more than 15 trillion tokens, the knowledge cutoff is still December 2023, the context window is 128K tokens, and grouped-query attention is used throughout. The eight officially supported languages are unchanged: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.

Practically, that means the serving profile is identical to Llama 3.1 70B. Anything that ran 3.1 70B, whether a vLLM deployment, a quantised build or a provider endpoint, runs 3.3 70B with the same memory and the same throughput. Hugging Face lists the checkpoint at about 71B parameters in BF16, so full-precision weights alone are on the order of 140 GB; a 4-bit build brings that to roughly a quarter, before the KV cache.

Where post-training moved the numbers

Meta's model card puts 3.3 70B beside both Llama 3.1 70B and Llama 3.1 405B, which is the comparison that matters:

• IFEval — 3.1 70B 87.5, 3.3 70B 92.1, 3.1 405B 88.6
• HumanEval (pass@1) — 80.5, 88.4, 89.0
• MATH (CoT) — 68.0, 77.0, 73.8
• MGSM — 86.9, 91.1, 91.6
• GPQA Diamond (CoT) — 48.0, 50.5, 49.0
• MMLU Pro (CoT) — 66.4, 68.9, 73.3

On instruction following and MATH, the 70B passes the 405B outright. On HumanEval and MGSM it effectively matches it. This is the result that made 3.3 70B popular: for many workloads it delivered 405B-class output at the serving cost of a 70B.

Where it did not move

Two rows in the same table are flat, and they are as informative as the gains.

MMLU with chain of thought is 86.0 for both 70B versions, while the 405B scores 88.6. MMLU Pro improves only modestly and stays well behind the 405B. Broad knowledge is largely fixed at pretraining, and a post-training pass did not change it. If your task depends on the model simply knowing more, the 405B still has an edge.

Function calling is also unchanged: BFCL v2 is 77.5 for 3.1 70B and 77.3 for 3.3 70B, against 81.1 for the 405B. Do not expect 3.3 to fix tool-selection problems you had with 3.1. If tool reliability is your bottleneck, the upgrade will not help on its own.

A drop-in replacement, with one caveat

Because the architecture and prompt format carry over, swapping 3.1 70B for 3.3 70B is usually a model id change. Meta's documentation notes support for several tool-use formats and Transformers chat templates, the same as 3.1.

The caveat is behaviour, not compatibility. Better instruction following means the model takes your prompt more literally. Prompts that were tuned to work around 3.1's habits, for example by repeating a format instruction three times, may produce different output. Run your existing eval before switching rather than after. If you are still on the older model, the Llama 3.1 70B guide covers what it does differently.

Llama 3.3 70B or Llama 4 Scout

The natural question in 2026 is whether to use 3.3 70B at all, given Llama 4. Meta's Llama 4 model card includes 3.3 70B as a reference point, which makes the comparison with Llama 4 Scout direct:

• MMLU Pro — 3.3 70B 68.9, Scout 74.3
• GPQA Diamond — 50.5, 57.2
• LiveCodeBench — 33.3, 32.8

Scout is ahead on knowledge and reasoning, roughly level on code, and adds image input and a much longer context window. But Scout is a 109B-total mixture-of-experts model, so it needs more memory than a 70B dense model even though it computes less per token. A quantised 70B can fit where Scout does not, and the dense model is simpler to serve and fine-tune. If you want the mechanics behind that trade, dense vs mixture-of-experts lays it out.

On price, OpenRouter lists 3.3 70B at $0.10 per million input tokens and $0.32 per million output at the time of writing, very close to what it lists for Scout. Through an API, then, the choice is about capability and modality rather than cost. Self-hosted, it is about memory.

When to use it

• You need solid instruction following and code at low cost, text only, with prompts under a few tens of thousands of tokens.
• You are fine-tuning, and want a dense base with mature tooling and a large community of existing adapters.
• You are self-hosting on hardware that can hold a quantised 70B but not a 109B MoE.
• You are replacing Llama 3.1 70B and want a free quality gain with no infrastructure change.

Pick something else when you need images, very long documents, stronger tool calling, or current knowledge. The weights are under the Llama 3.3 Community License, with the familiar 700 million monthly active user threshold and "Built with Llama" attribution. Before committing, run your own tasks on it: a twenty-task eval will tell you more than the table above.