Meta AI's Llama 3.3 70B Instruct is an open-source, text-in, text-out chat model. It is designed for tasks involving conversation, code execution, web search via integrated tools, and structured output generation.
Input
Output
Context
131K
Max Output
4K
Parameters
70.6B
Input Modalities
Output Modalities
Some values come from web research and may be incorrect. Report incorrect data
Measured speed: 97 tok/s output speed · 0.66 s to first answer token · 0.66 s to first token
Bands compare this model with every model the same suite measured in our catalog.
Download 141 GB · Runs on vLLM v0.32.0 · 46 GGUF builds
| Precision | Weights | One request at 33K tokens | Smallest GPU that fits |
|---|---|---|---|
| FP16 / BF16 | 131 GB | 147 GB | 1x NVIDIA B200 SXM (192 GB) |
| INT8 | 65.7 GB | 77.7 GB | 1x NVIDIA H100 NVL (94 GB) |
| INT4 | 32.9 GB | 43.2 GB | 1x Biren BR100 (64 GB) |
Computed by Inferbase from the 70,553,706,496 parameters in the creator's Hugging Face repository; one request adds its KV cache, activations and framework overhead, with the layer layout estimated from the parameter count.
Request meta-llama-3-3-70b-instruct and routing picks a build. Use a build's model ID to pin it.
| Build | What it is | Model ID |
|---|---|---|
| Turbo | Optimized serving build | meta-llama-3-3-70b-instruct:turbo |
Public resources
Specification sources
As of Nov 10, 2026.
Answered from the facts on this page, with their sources.
Llama 3.3 70B Instruct has a context window of 131,072 tokens and a maximum output of 4,000 tokens (source: Creator or provider API). A provider may serve less; Inferbase routes on what each provider serves.
Yes. Llama 3.3 70B Instruct is an open-weight model under the Llama3.3 license. You can download and self-host it within the license terms.
Llama 3.3 70B Instruct has 70,553,706,496 parameters (70.6B), as stated in the creator's Hugging Face repository.
Strengths: 131K token context window. Limitations: Below average on graduate-level science (GPQA) among measured models; Below average on knowledge (MMLU-Pro) among measured models; Below average on competition math (AIME 2025) among measured models. Each is computed from the model's stated facts and benchmark results.
Llama 3.3 70B Instruct accepts Text and produces Text. The Capabilities card lists the API features it supports, such as function calling and structured output.
Llama 3.3 70B Instruct has published results from Artificial Analysis and Epoch AI. The Benchmarks card shows each score as reported, with its band among the models the same source measured. Scores are never combined across sources.
Yes. Llama 3.3 70B Instruct is available on Inferbase Managed through the OpenAI-compatible API as meta-llama-3-3-70b-instruct. Request it by name, or request auto and routing picks it when it fits the request best.
The weights take 131.4 GB at FP16 / BF16, 65.7 GB at INT8, 32.9 GB at INT4, computed from the 70,553,706,496 parameters its creator publishes. Each request adds its KV cache and the framework's overhead. The hardware section shows one request and the smallest GPU that fits; the GPU planner sizes it for your context length and traffic.
The model card of Llama 3.3 70B Instruct lists 8 languages: English, French, Italian, Portuguese, Hindi, Spanish, Thai and German. That is what Meta AI states; the catalog does not test language quality.
Start on the free plan with 5,000 requests a month, no card required.