Meta Llama 3.3
by Meta Platforms
70B open-weight model with 405B-class quality at a fraction of the serving cost
Llama 3.3 is Meta's 70B open-weight, instruction-tuned language model, released in December 2024. It delivers accuracy close to the far larger 405B Llama 3.1 at a fraction of the inference cost, giving enterprises a self-hostable multilingual model they can run on their own GPUs, fine-tune freely, and deploy without per-token API fees or hard vendor lock-in.
Llama 3.3 is Meta's 70-billion-parameter instruction-tuned text model, released on 6 December 2024 as the final refresh of the Llama 3 series. It is a dense decoder-only transformer trained on more than 15 trillion tokens with a December 2023 knowledge cutoff, using grouped-query attention and a 128,000-token context window, aligned with supervised fine-tuning plus reinforcement learning from human feedback. Meta's own model card puts it within a few points of the far larger Llama 3.1 405B on general reasoning - 86.0 versus 88.6 on MMLU with chain-of-thought, 88.4 versus 89.0 on HumanEval - while beating it on MATH at 77.0 versus 73.8. That gap is the entire commercial argument for the model: roughly 405B-class quality at a fraction of the serving cost. Training consumed 7.0 million H100-80GB GPU hours. It officially supports eight languages (English, German, French, Italian, Portuguese, Hindi, Spanish and Thai), emits JSON for function calling, and is distributed as open weights under the Llama 3.3 Community License rather than a true open-source licence - organisations with more than 700 million monthly active users must negotiate a separate agreement with Meta, and the acceptable-use policy bars military, weapons and surveillance applications. Enterprises typically consume it through a host rather than raw weights: it shipped day-one on IBM watsonx.ai, SambaNova Cloud and NVIDIA's NIM catalogue, and runs locally via Ollama, llama.cpp, LM Studio and vLLM. Quantised to 4-bit it fits a single 48GB workstation GPU, which is what made it a default self-hosted enterprise baseline through 2025.
Platform and ML engineering teams that need a capable general-purpose LLM running inside their own VPC or data centre, where data cannot leave the perimeter and per-token API bills are the thing being escaped.
Near-405B reasoning quality on hardware you already own, with no per-token fee and no requirement to send prompts to a third-party API.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Free, Usage-based
- Target Market
- CTOs, Heads of ML Platform, Enterprise Developers, Data Scientists, AI Infrastructure Leads
- Deployment
- Open-source, Self-hosted, API-based, Hybrid
- Founded
- 2004
- Headquarters
- Menlo Park, United States
- Team Size
- 500+
Key Features
- ✓70B dense weights with 405B-adjacent quality
Scores 86.0 MMLU CoT against Llama 3.1 405B's 88.6 while needing roughly a sixth of the serving hardware, which is what makes local deployment economically viable.
- ✓128,000-token context window
Holds long contracts, codebases or multi-document retrieval sets in a single prompt without an external chunking and re-ranking layer.
- ✓Structured JSON output and function calling
Generates well-formed JSON for tool invocation, which is the prerequisite for using it as the reasoning core of an agent rather than a chatbot.
- ✓Eight officially supported languages
English, German, French, Italian, Portuguese, Hindi, Spanish and Thai are trained and evaluated, so multilingual support is measured rather than incidental.
- ✓Open weights with distillation rights
The licence permits using outputs to train and improve other models, so teams can distil a small task-specific model from it legally.
- ✓Broad runtime support out of the box
Ollama, llama.cpp, vLLM, LM Studio, watsonx.ai, SambaNova Cloud and NVIDIA NIM all serve it, so the deployment path is not a research project.
Capabilities
Use Cases
- •Air-gapped internal assistant
Defence, health and financial teams run the weights inside their own network so no prompt or document ever reaches an external API endpoint.
- •Synthetic training data generation
Generate labelled examples at volume to fine-tune a small specialist model, using the licence's explicit permission to train on Llama outputs.
- •Cost-capped high-volume classification
Route millions of routine ticket or document classifications to owned GPUs, converting an unbounded per-token API bill into a fixed hardware cost.
- •Multilingual support triage
Summarise and route inbound tickets across eight supported languages without paying a separate translation vendor for each queue.
- •On-device and workstation prototyping
Quantised to 4-bit the model fits a 48GB workstation GPU, letting engineers iterate on prompts offline before committing to cluster capacity.
Ideal For
Best For
- ✓Self-hosted enterprise chat assistants where prompt data must stay inside the network perimeter
- ✓Synthetic data generation and distillation into smaller task-specific models, explicitly permitted by the licence
- ✓Multilingual customer support across the eight officially supported languages
- ✓Tool-calling and agentic backends that need reliable JSON function-call output without API metering
- ✓Fine-tuning on proprietary domain data where weight ownership matters more than frontier benchmark scores
Not Ideal For
- ✗Consumer-scale products above 700 million monthly active users - the Llama 3.3 Community License requires a separately negotiated agreement with Meta at that threshold
- ✗Teams without GPU capacity or MLOps staff; running 70B weights well needs 48GB+ of VRAM and quantisation expertise, and Hacker News practitioners report only 1-3 tokens/sec on a single consumer 4090 at Q4
- ✗Vision, audio or long-horizon reasoning workloads - Llama 3.3 is text-in/text-out with no built-in reasoning mode, so a multimodal or reasoning-tuned model is the better fit
- ✗Buyers who need a vendor-backed SLA, indemnity and compliance attestations; the weights ship as-is and every certification comes from whichever host you run them on
Integrations
Deployment
Market Analysis
Pros
- ✓Roughly 405B-class reasoning quality at 70B serving cost - 86.0 vs 88.6 MMLU CoT, and it actually beats the 405B on MATH
- ✓Genuinely portable: the same weights run on Ollama locally, vLLM in your cluster, or watsonx.ai and Bedrock as a managed endpoint
- ✓128k context and native JSON function calling make it usable as an agent backend, not just a chat model
- ✓Zero licence cost below 700M MAU, which removes per-token API spend from high-volume classification and summarisation workloads
Cons
- ✗Called open source but is not - Hacker News commenters on the release thread pushed back hard, noting the weights cannot be rebuilt from scratch and the 700M MAU clause plus acceptable-use policy make it open-weight freeware rather than OSI-licensed software
- ✗Practical inference speed disappoints on consumer hardware: practitioners in that thread reported around 2 tokens/sec for Q4_K_M on a single RTX 4090, with 48GB of VRAM being the realistic floor for usable throughput
- ✗Quantisation is not free - the same discussion reported measurable degradation in strict instruction-following and factual accuracy at lower bit depths, which is exactly where the cost savings come from
- ✗Text-only with a December 2023 knowledge cutoff and no reasoning mode, so it is behind newer multimodal and reasoning-tuned releases on anything current or visual
- ✗Smaller competitors have narrowed the gap; the same practitioners repeatedly named Qwen 2.5 32B as the better fit for consumer hardware
Pricing
Llama 3.3 Community License (open weights)
$0
- ✓Full 70B instruction-tuned weights
- ✓Commercial use permitted below 700M MAU
- ✓Fine-tuning and distillation rights
- ✓No per-token fee when self-hosted
Above 700M monthly active users
Contact for pricing
- ✓Separate licence negotiated directly with Meta
- ✓Required before the model may be used at that scale
The weights themselves cost nothing under the Llama 3.3 Community License, so the real bill is either GPU capacity you own or the per-token rate charged by whichever host serves it - IBM watsonx.ai, SambaNova Cloud, NVIDIA NIM, Groq, Together and Bedrock all price it independently and none of that money goes to Meta. The one hard commercial gate is the 700 million monthly-active-user threshold, above which a separate agreement with Meta is mandatory; below it, commercial use is free but subject to an acceptable-use policy that bars military, weapons and surveillance applications.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 5 on domains other than llama.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.