Llama 3.3 70B

What's New in Llama 3.3 vs 3.1

Llama 3.3 70B, released in December 2024, is Meta's instruction-tuned update to the 70B. It keeps Llama 3.1 70B's architecture, parameter count and 128K context, so it is a drop-in replacement on the same hardware; the gains come from post-training. All figures below are from Meta's Llama 3.3 model card, which reports both models side by side with the same settings.

Two rows on the card move the other way or not at all: MMLU is level, and tool use (BFCL v2) is 77.5 for 3.1 and 77.3 for 3.3. Same hardware requirements as Llama 3.1 70B, so it is a direct upgrade. Scores are Meta's, from Meta's evaluation setup; this site has not run them.

Llama 3 to 3.1 to 3.3: the three 70B releases

Llama 3 70B (April 2024), Llama 3.1 70B (July 2024) and Llama 3.3 70B (December 2024) are the same 70.55-billion-parameter network shape: 80 layers, hidden size 8192, 64 attention heads sharing 8 key/value heads. What changed is the context window (8K for Llama 3, 128K from 3.1 on), the training and post-training recipe, and the licence, which Meta reissues with each version. If you searched for “Llama 3 70B size”, the answer is 40.0 GB for Ollama's llama3:70b (q4_0) and 42.5 GB at Q4_K_M; the newer releases are the same 42.5 GB at Q4_K_M.

Sources: the Llama 3 model card, Llama 3.1 card and Llama 3.3 card; sizes from the Ollama registry, 2 November 2026. The Llama 3 card used different shot counts (5-shot MMLU, 4-shot MATH), so its column is not directly comparable with the other two.

In practice there is no reason to pull llama3:70b for new work: it needs the same hardware as Llama 3.3 and has a sixteenth of the context. The Llama 3 licence also carries naming rules (derived models must start with “Llama 3”) and a 700-million-monthly-user threshold, which the later licences keep. For the earlier 70B generation, the Llama 2 70B spec page covers its 4K-context architecture, and Llama 3.1 70B has its own page.

Technical Specifications

Benchmark Performance

Benchmark scores from Meta's official model card. MMLU = general knowledge (0-shot, chain of thought), MMLU Pro = a harder 10-option version, IFEval = instruction following, HumanEval = Python coding, MATH = competition maths, MGSM = grade-school maths in ten languages. On the same card Llama 3.1 405B scores 88.6 (MMLU), 73.3 (MMLU Pro) and 89.0 (HumanEval), so the 70B gets most of the way to a model almost six times its size. No speed figures appear here because none were measured; the hardware section gives bandwidth ceilings instead.

Quick Start: Run Llama 3.3 70B with Ollama

The layer list above is the real manifest for llama3.3:70b: one 42.5 GB model layer, a chat template, two licence files and a parameters file.

Hardware Requirements

The 42.5 GB Q4_K_M file needs about 46 GB with an 8K context: 42.5 GB of weights, 2.7 GB of KV cache (327,680 bytes per token at f16) and roughly 1 GB of compute buffers. Three setups hold that comfortably at home: two 24 GB cards, one 48 GB workstation card, or an Apple Silicon Mac with 64 GB of unified memory. Everything else is either a lower-bit file or a CPU/GPU split.

VRAM by Quantization

Every speed in this table is a ceiling, not a measurement: the vendor's memory bandwidth divided by the file size, because generating each token reads every weight once. When a model is split across two cards the layers run one card at a time, so the ceiling is one card's bandwidth over the whole file, not the sum. Real runs land below these numbers; on the one machine this site has measured (an Apple M3 Pro), 7-8B models reached about 80% of their ceiling. For other cards, the 24 GB and 32 GB VRAM guides list what fits a single card, and the Mac memory pressure guide explains why a 64 GB Mac is the practical floor for a 70B at 4-bit.

Llama 3.3 70B vs Competitors

Scores are each developer's own, from different evaluation setups, so read them as rough positioning. Speeds are bandwidth ceilings on a pair of RTX 4090s (1,008 GB/s ÷ file size). Mixtral 8x22B is a mixture-of-experts model, so file size does not set its speed and no ceiling is quoted.

Best Use Cases for Llama 3.3 70B

Quantization Guide

Ollama's default llama3.3:70b is Q4_K_M, which is the usual balance of quality and memory for a 70B. Every tag below exists on the Ollama registry; sizes are its download sizes, and the memory column adds 2.7 GB of KV cache for an 8K context plus about 1 GB of buffers.

The quality column is the general trade-off of each quantisation level, not a measured score for this model; Meta publishes no figures for quantised builds. A q4_0 file (40.0 GB, llama3.3:70b-instruct-q4_0) and Q6_K (57.9 GB) also exist. For a detailed explanation of quantization formats, see quantization explained.

Advanced Setup

Custom Modelfile

Create a tuned version with custom parameters. A 16K context adds about 5.4 GB of KV cache, so keep it at 8192 on a 48 GB setup if ollama ps shows a CPU split:

API Access

Use Ollama's OpenAI-compatible API:

Pair with a Web Interface

For the best experience, pair Llama 3.3 70B with Open WebUI for a full ChatGPT-like interface, or use it as the backend for Continue.dev for AI-assisted coding.

Frequently Asked Questions

Sources