Model Details
Llama 3.1 405B Instruct
Quick answer: Llama 3.1 405B Instruct is Meta's July 2024 largest open-weight model, scoring 96.9% on AI2 Reasoning Challenge, 88.6% on MMLU, 73.3% on MMLU-Pro, and 26.4% on BigCodeBench. With a 128K-token context window under the Llama 3.1 License, it was the first open-weight model to match GPT-4-class performance at the 400B+ parameter scale.
At a Glance
Where Llama 3.1 405B leads
- 96.9% AI2 Reasoning Challenge — near-perfect scientific reasoning
- 88.6% MMLU — matches GPT-4 Turbo (86.4%) at open-weight pricing
- 73.3% MMLU-Pro — strong academic knowledge
- Largest open-weight dense model at launch
- 128K context window
- Establishes frontier capability ceiling for open-weight models at launch
Where it lags
- Dense 405B: massive infrastructure requirement for self-hosting
- Superseded by Llama 4 Maverick for most tasks at far lower cost
- Text-only: no vision
- March 2024 knowledge cutoff
- 26.4% BigCodeBench — moderate coding performance
Best for: Research requiring maximum open-weight capability; self-hosted frontier AI where data sovereignty is paramount; fine-tuning at maximum scale.
What Llama 3.1 405B Instruct Is
Llama 3.1 405B Instruct was the largest open-weight model available when released July 23, 2024 — and the first to directly compete with GPT-4 on MMLU. Its release was a landmark event: for the first time, teams could access GPT-4-class performance through self-hosted open weights.
At 88.6% MMLU, the 405B surpassed the then-state-of-the-art GPT-4 Turbo (86.4%) on the primary academic knowledge benchmark. This performance came at substantial infrastructure cost: a dense 405B model requires multiple high-end GPUs for inference. For most production use cases, Llama 4 Maverick (17B active MoE, 1M context, better overall performance) is the recommended replacement.
Llama 3.1 405B remains relevant for teams with specific requirements: maximum open-weight capability, dense architecture for quantisation research, or use cases where the Llama 3.1 instruction style has been validated in production.
Specifications
| Field | Value |
|---|---|
| Organization | Meta |
| Parameters | 405B (dense) |
| Context window | 128,000 tokens |
| License | Llama 3.1 License |
| HuggingFace | meta-llama/Meta-Llama-3.1-405B-Instruct |
| Release date | July 23, 2024 |
| Knowledge cutoff | March 2024 |
| Modality | Text only |
Pricing
Llama 3.1 405B is available as open weights under the Llama 3.1 License — free to self-host. Hosted API access via Together AI, Fireworks, and others at market rates.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| AI2 Reasoning Challenge | 96.9% | Benchgen evaluation | 2024-07 |
| MMLU | 88.6% | Benchgen evaluation | 2024-07 |
| MMLU-Pro | 73.3% | Benchgen evaluation | 2024-07 |
| BigCodeBench | 26.4% | Benchgen evaluation | 2024-07 |
Llama 3.1 405B vs Alternatives
| Model | MMLU | MMLU-Pro | Context | Weights |
|---|---|---|---|---|
| Llama 3.1 405B Instruct | 88.6% | 73.3% | 128K | Open (dense 405B) |
| Llama 3.1 70B Instruct | 86.0% | — | 128K | Open (dense 70B) |
| Llama 4 Maverick | — | 80.5% | 1M | Open (17B active MoE) |
| GPT-4 Turbo | 86.4% | — | 128K | Closed |
Llama 4 Maverick (17B active MoE) achieves 80.5% MMLU-Pro vs Llama 3.1 405B's 73.3%, with a 1M context window, multimodal capability, and dramatically lower inference cost. For new deployments, Llama 4 Maverick is the recommended open-weight choice.
Related Models & Benchmarks
- Compare: Llama 3.1 70B Instruct · Llama 4 Maverick · DeepSeek-V3
- Benchmarks: AI2 Reasoning Challenge · MMLU · MMLU-Pro
Frequently Asked Questions
What is Llama 3.1 405B?
Llama 3.1 405B Instruct is Meta's July 2024 largest open-weight model, scoring 96.9% AI2 RC, 88.6% MMLU, and 73.3% MMLU-Pro with 128K context. It was the first open-weight model to match GPT-4-class MMLU performance.Is Llama 3.1 405B open source?
The weights are open on Hugging Face under the Llama 3.1 License. Commercial use is permitted with some restrictions.Should I use Llama 3.1 405B or Llama 4 Maverick?
For new deployments, Llama 4 Maverick is strongly preferred: better benchmark scores, 1M context, multimodal capability, and dramatically lower inference cost (17B active MoE vs 405B dense). Llama 3.1 405B is appropriate for existing integrations or when a dense 405B architecture is specifically required.Specs from Meta's official Llama 3.1 announcement (July 2024) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.