Is Llama actually free?

The weights are openly licensed, but running them is not free. You either pay a hosting provider per token or pay for your own GPU capacity, engineering time and idle headroom.

When does self-hosting Llama beat an API?

Typically at sustained high volume with predictable load. At low or spiky volume, per-token hosted pricing usually wins because you are not paying for idle GPUs.

How do I compare Llama to a closed model?

Price the same workload both ways: hosted per-token rate versus your all-in infrastructure cost per token, including utilisation.

Meta Llama Pricing: Hosted Rates and Self-Hosting Cost

Hosted per-token rates

Llama 3.3 70B at Together/Fireworks: ~$0.88/1M blended. Groq: ~$0.59/$0.79 with 5× the throughput. DeepInfra often undercuts both for batch workloads.

Self-hosting cost model

A single H100 sustains ~2K tokens/sec on 70B INT8. At $2/hr on-demand, that's ~$0.28/1M tokens — cheaper than hosted only above ~60% utilization.

Break-even with hosted APIs

Self-hosting wins at >5M tokens/day sustained. Below that, hosted is cheaper after factoring ops, autoscaling gaps, and engineer time.

Latency-priced providers

Groq and Cerebras charge similar $/token but deliver 5–10× throughput on Llama. Use when latency drives revenue (voice, real-time agents).

When Llama beats closed models

Privacy-sensitive workloads, regulated environments, fine-tune-heavy use cases, or sustained high-volume routes where price per token dominates.

Worked example: hosted vs self-hosted break-even

At Groq's $0.59/$0.79 blended rate on 70/30 input/output, a 10M-token/day workload costs roughly 10M*0.7*0.59/1e6 + 10M*0.3*0.79/1e6 ≈ $4.13+$2.37=$6.50/day, or about $195/month. A single on-demand H100 at $2/hr running 24/7 costs $1,440/month regardless of how many tokens it processes — self-hosting only wins once sustained volume is high enough to keep that GPU near-saturated most hours.

Mechanics: why utilization drives the self-host math

A GPU bills by the hour whether it's processing tokens or idle, while hosted APIs bill only for tokens actually sent. The $0.28/1M self-hosted estimate assumes near-continuous saturation; at 20% utilization the effective per-token cost is roughly five times higher because the same hourly cost is spread over far fewer tokens. This is the single biggest variable teams underestimate when comparing the two paths.

Decision checklist: hosted API vs self-hosting

— Is sustained daily volume above roughly 5M tokens/day, the rough break-even point? — Do you have the ops capacity to manage autoscaling, failover, and model updates? — Does the workload have latency requirements that favor a dedicated GPU over shared hosted capacity? — Have you priced at least two hosted providers (Groq, Together, Fireworks) before assuming self-hosting is cheaper?

What this pricing comparison does not cover

This post does not cover fine-tuning cost, multi-GPU serving architectures, or quantization trade-offs (INT8 vs FP16) in detail — those materially change the self-hosting math and deserve dedicated modelling once you're past the initial hosted-vs-self-host decision covered here.

Frequently asked questions

Is Llama actually free?: The weights are openly licensed, but running them is not free. You either pay a hosting provider per token or pay for your own GPU capacity, engineering time and idle headroom.
When does self-hosting Llama beat an API?: Typically at sustained high volume with predictable load. At low or spiky volume, per-token hosted pricing usually wins because you are not paying for idle GPUs.
How do I compare Llama to a closed model?: Price the same workload both ways: hosted per-token rate versus your all-in infrastructure cost per token, including utilisation.