Llama 3.1 70B Instruct vs Llama 3.1 8B Instruct
| Spec | Llama 3.1 70B Instruct | Llama 3.1 8B Instruct |
|---|---|---|
| Respan Index | 10.0 (#159) | 6.9 (#161) |
| Blended price per 1M | $0.4 | $0.025 |
| Context window | 128K | 128K |
| Input / output per 1M | $0.4 / $0.4 | $0.02 / $0.04 |
| Max output | - | - |
| Input | text | text |
| Released | 2024-07-23 | 2024-07-23 |
Respan Index by category
Each category is scored 0 to 100, and how far ahead the winner is. Leads under a point are close to even.
| Category | Llama 3.1 70B Instruct | Llama 3.1 8B Instruct |
|---|---|---|
| Coding | 0.4 Llama 3.1 70B Instruct | 0.0 Llama 3.1 8B Instruct |
| Agents | 4.9 Llama 3.1 70B Instruct | 0.8 Llama 3.1 8B Instruct |
| Reasoning | 15.3 Llama 3.1 70B Instruct | 13.8 Llama 3.1 8B Instruct |
| Knowledge & language | 2.1 Llama 3.1 70B Instruct | 0.4 Llama 3.1 8B Instruct |
| Human preference | 30.6 Llama 3.1 70B Instruct | 21.6 Llama 3.1 8B Instruct |
Benchmarks both report
Llama 3.1 70B Instruct leads on 11 of the 11 benchmarks both report. Higher is better. Each score links to where it was published.
| Benchmark | Llama 3.1 70B Instruct | Llama 3.1 8B Instruct |
|---|---|---|
| GPQA Diamond | 44.2% | 27% |
| MATH Level 5 (Epoch AI) | 36.7% | 22.9% |
| OTIS Mock AIME 2024-2025 (Epoch AI) | 3.6% | 1.7% |
| WeirdML (Håvard Tveit Ihle) | 9% | 1.7% |
| LMArena Elo | 1293 | 1211 |
| BALROG (BALROG) | 27.9% | 15.1% |
| BFCL | 84.8% | 76.1% |
| HumanEval | 80.5% | 72.6% |
| IFEval | 87.5% | 80.4% |
| MMLU | 83.6% | 69.4% |
| MMLU-Pro | 66.4% | 48.3% |
FAQ
Is Llama 3.1 70B Instruct better than Llama 3.1 8B Instruct?
Llama 3.1 70B Instruct scores 10.0 on the Respan Index (#159), ahead of Llama 3.1 8B Instruct at 6.9 (#161). Llama 3.1 70B Instruct leads on 11 of the 11 benchmarks both report.
Which is cheaper, Llama 3.1 70B Instruct or Llama 3.1 8B Instruct?
Llama 3.1 70B Instruct lists at $0.4 input and $0.4 output per 1M tokens; Llama 3.1 8B Instruct at $0.02 and $0.04. Llama 3.1 8B Instruct is 16x cheaper per token at list price.
Which has the larger context window?
Both accept 128K tokens of context.