DEPLOY

Open-source LLM · Meta

Llama 3.3 70B

70 B parameters · 128K context · Llama 3.3 Community License · released 2024-12-06 by Meta.

Llama 3.3 70B needs about 140 GB at FP16: no single GPU can hold it. Serving requires an 8-GPU HGX baseboard (INT8/INT4) or a full rack (NVL72, TPU pod) at FP16.

Where this model sits

Meta's post-training refresh of the 70B: matches Llama 3.1 405B quality on many benchmarks at a fraction of the serving cost.

What runs it (production serving)

Datacenter GPUs and rack designs sized for Llama 3.3 70B. Consumer cards are not practical at this size.

Memory at different context lengths

Weights are fixed. The KV cache grows linearly with context length and can rival the weights at full window. Most guides only quote weights, then hand-wave a 15% overhead, which is only true at short context.

ContextFP16 totalINT8 totalINT4 total
4K (short chat)140 + 1.3 = 141.3 GB70 + 0.6 = 70.6 GB35 + 0.6 = 35.6 GB
32K (typical RAG)140 + 10.0 = 150.0 GB70 + 5.0 = 75.0 GB35 + 5.0 = 40.0 GB
128K (full window)140 + 40.0 = 180.0 GB70 + 20.0 = 90.0 GB35 + 20.0 = 55.0 GB

KV bytes per token = 2 × 80 layers × 8 KV heads × 128 head-dim × precision bytes. INT4 rows keep the KV cache at INT8 (typical serving default).

Quality benchmarks

Vendor-reported scores from the release blog or tech report. Every row cites its source.

BenchmarkScoreSourceDate
MMLU (5-shot)86.0Meta Llama 3.3 model card2024-12-06
HumanEval88.4Meta Llama 3.3 model card2024-12-06
MATH (0-shot)77.0Meta Llama 3.3 model card2024-12-06
GPQA50.5Meta Llama 3.3 model card2024-12-06

Speed and cost to serve

What cloud providers currently charge to serve Llama 3.3 70B. Tokens per second is what the provider publishes on their own page. Batch size, prompt length, and inference engine can shift these numbers a lot; treat as directional.

ProviderHardwareSpeed$/M in$/M outAs of
Together AI
Instruct Turbo
H100—$0.88$0.882025-08-01
GroqGroq LPU250 tok/s$0.59$0.792025-08-01
Fireworks AI——$0.90$0.902025-08-01

Cheapest option in this list for output-heavy workloads: Groq at $0.79/M out. Fastest published: Groq at 250 tok/s (Groq LPU).

License and variants

License
Llama 3.3 Community License
The gotcha
700M-MAU commercial threshold, same as prior Llama Community License.
Variants
Instruct
Context window
131,072 tokens (128K)
Architecture
dense

Compare with

Model card: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct · Release: Meta announcement

See also: every open-source LLM we track · all AI chips · rack designs.