Llama
Meta's open-weight family, and the reason local models took off.
Llama is Meta's open-weight family and the one that made running a serious model on your own hardware normal — most local tooling was built for it first, so support is everywhere. The Llama 4 models are mixture-of-experts and multimodal, with very large context windows, and the line includes Llama Guard, a classifier for filtering input and output rather than a chat model. Meta's newer work ships under a different name, so check the dates on this family before choosing it.
At a glance — Llama 3.1 8B Instruct
- Cost
- $0.05 in · $0.08 out per 1M tokensBudget
- Memory
- 131K tokens — about 200 pages
- Reads
- Text
- Can do
- Use tools (agents)
- Structured output
- Thinking mode (not supported)
- Good at
- Coding: better than 1%
- Reliability
- 99.99% uptime, last 24h
- Released
- Jul 23, 2024
Advanced details
- API model ID
meta-llama/llama-3.1-8b-instruct- Max output
- 118K tokens
- Cached input
- $0.025 per 1M
- Knowledge cutoff
- 2023-12-31
- Tokenizer
- Llama3
- Coding index
- 5.4
Providers
The same model, hosted by different companies. OpenRouter picks one per request and falls back to the next if it fails.
| Provider | Input /1M | Output /1M | Cached /1M | Max output | Uptime 24h |
|---|---|---|---|---|---|
| DeepInfra · fp8 | $0.02 | $0.04 | — | 16K | 99.77% |
| Novita · fp8 | $0.02 | $0.05 | — | 15K | 97.52% |
| Groq | $0.05 | $0.08 | $0.025 | 118K | 98.90% |
| Cloudflare · fp8 | $0.152 | $0.287 | — | 29K | 98.86% |
| CoreWeave · bf16 | $0.22 | $0.22 | $0.22 | 118K | 99.99% |
Supported parameters
- frequency_penalty
- logit_bias
- logprobs
- max_tokens
- min_p
- presence_penalty
- repetition_penalty
- response_format
- seed
- stop
- structured_outputs
- temperature
- tool_choice
- tools
- top_k
- top_logprobs
- top_p
Live from OpenRouter, refreshed hourly. Scores from Artificial Analysis.
Key info
- Pricing
- Open source
- Category
- Models
- Best for
- Intermediate
- Made by
- Meta
- Weights
- Open
- Good for
- Local setups, where tooling support is widest