Llama 3.3 70B Instruct

Deploy Llama 3.3 70B Instruct on a dedicated GPU from $0.55 per hour, billed per minute, on 1× RTX A6000. Meta's 70B multilingual instruction model with 128K context, strong code and tool-use results, and substantially improved instruction following over Llama 3.1 70B.

Text70B paramsLlama 3.3 Community LicenseDeploys in ~5 minSecure Cloud

meta-llama/Llama-3.3-70B-Instruct

HexGrid Cloud price

$0.55/hr

1× RTX A6000

Context

128K

tokens

Parameters

70B

dense

Languages

officially supported

Running Llama 3.3 70B Instruct on HexGrid Cloud

What you get when you deploy with us, beyond the hourly rate.

Single tenant by default

The model runs on a GPU that is yours for the life of the instance. No shared endpoint, no queue behind other tenants.

Per-minute billing

You pay for the GPU, not the token. Stop the instance and billing stops with it.

OpenAI-compatible endpoint

Point an existing SDK at your instance by changing the base URL. No rewrite required.

Your weights, your data

Prompts and outputs stay inside your instance. Nothing is logged, sampled or used for training.

Llama 3.3 70B Instruct GPU sizing and cost

What the model needs at each precision, and the cheapest HexGrid Cloud configuration that holds it.

Llama 3.3 70B Instruct vRAM requirement and hourly cost by precision
Precision	vRAM needed	Runs on	Price
BF16 Recommended	168 GB	4× RTX A6000	$2.20/hr
FP8	84 GB	2× RTX A6000	$1.10/hr
INT4	42 GB	1× RTX A6000	$0.55/hr

No universal tokens-per-second figure should be published without specifying GPU, tensor parallelism, batch size, context length and serving engine. Native examples use BF16.

About Llama 3.3 70B Instruct

What Llama 3.3 70B Instruct is built for, and where it falls short.

Llama 3.3 70B Instruct is Meta's text-only, instruction-tuned 70B Llama model. It uses an autoregressive Transformer with grouped-query attention and was post-trained with supervised fine-tuning and reinforcement learning with human feedback.

The model supports a 128K context window and is optimized for multilingual dialogue, coding, instruction following and tool use. Meta officially lists English, German, French, Italian, Portuguese, Hindi, Spanish and Thai as supported languages.

It is served through the same OpenAI-compatible chat-completions pattern used by common inference engines such as vLLM and SGLang.

What people run it for

General AI assistants

High-quality conversational assistants requiring strong instruction following, writing and reasoning.

Coding assistants

Code generation, explanation, debugging and tool-assisted engineering workflows.

Tool-using agents

Agents that call structured functions, APIs and other application tools.

Multilingual applications

Dialogue and generation in Meta's eight officially supported languages.

Strengths

Strong instruction following, with 92.1 on Meta's IFEval evaluation
High code-generation performance for the Llama 3 generation, including 88.4 HumanEval pass@1
128K context for long conversations, documents and agent state
Native tool-use formatting supported by the model's chat templates
Eight officially supported spoken languages plus code output

Limitations

70B dense weights make the model significantly more expensive to host than 8B-30B alternatives
Text-only model; it cannot directly consume images, audio or video
Knowledge cutoff is December 2023 according to Meta's model card
The Llama 3.3 Community License contains terms beyond a standard permissive open-source licence, including additional provisions for very large commercial services
128K context substantially increases KV-cache memory requirements when fully utilized

Quickstart

Serve with an OpenAI-compatible engine such as vLLM, then point an OpenAI client at the instance.

pythonPOST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-instance>.hexgrid.cloud/v1",
    api_key="<your-key>",
)

response = client.chat.completions.create(
    model="meta-llama/Llama-3.3-70B-Instruct",
    messages=[
        {"role": "system", "content": "You are a helpful technical assistant."},
        {"role": "user", "content": "Explain how grouped-query attention reduces serving cost."}
    ],
)

print(response.choices[0].message.content)

Meta's Hugging Face model page documents serving through vLLM and SGLang with an OpenAI-compatible /v1/chat/completions endpoint.

Llama 3.3 70B Instruct reported benchmarks

Scores published by the model authors. A starting point, not a guarantee on your own data.

MMLU CoT86
MMLU-Pro CoT68.9
IFEval92.1
GPQA Diamond CoT50.5
HumanEval pass@188.4
MATH CoT77
BFCL v277.3

Llama 3.3 70B Instruct specifications

Architecture and serving details for Llama 3.3 70B Instruct.

PublisherMetaParameters70BLicenceLlama 3.3 Community LicenseReleasedDecember 2024

Architecture

Type: Dense autoregressive Transformer
Parameters: 70B
Attention: Grouped-query attention (GQA)
Post-training: SFT + RLHF

Context

Context window: 128K tokens
Input: Multilingual text
Output: Text and code
Knowledge cutoff: December 2023

Serving

API: OpenAI-compatible chat completions
Engines: vLLM, SGLang, Transformers
Transformers: >= 4.45.0 per Meta model card
Tool use: Supported

Llama 3.3 70B Instruct frequently asked questions

Common questions about deploying Llama 3.3 70B Instruct on HexGrid Cloud.

Is Llama 3.3 70B Apache 2.0?

No. It is distributed under Meta's Llama 3.3 Community License, which has its own redistribution, attribution, acceptable-use and large-service commercial provisions.

Does Llama 3.3 70B support images?

No. This checkpoint is text-only.

What is its context window?

Meta lists a 128K-token context window.

Does it support tool calling?

Yes. Meta documents tool-use formats and Transformers chat-template support for structured tools.

Deploy Llama 3.3 70B Instruct today

$0.55 per hour on 1× RTX A6000, billed per minute, never shared.

Deploy Llama 3.3 70B Instruct See all models