llama-4-scout-17b-16e

llama-4-scout-17b-16e is a efficient-tier language model from Cerebras with 17B parameters. llama-4-scout-17b-16e is an open-weights model, meaning the weights are publicly available for self-hosting, fine-tuning, and deployment on your own infrastructure — not just through Cerebras's API. Llama 4 is Meta's latest open-weights generation, featuring a mixture-of-experts architecture and native multimodal support — a significant architectural shift from the dense Llama 3 family.

llama-4-scout-17b-16e is priced at $0.1/1M input tokens and $0.1/1M output tokens. Output tokens are 1× more expensive than input tokens — this ratio reflects the sequential nature of autoregressive generation, which requires a separate forward pass per output token versus parallel processing of all input tokens simultaneously. At a typical usage pattern of 1,000 input tokens and 500 output tokens per call, llama-4-scout-17b-16e costs approximately $0.15 per 1,000 API calls. Cerebras does not publish a prompt caching rate for this model, so repeated long contexts are billed at the standard $0.1/1M input rate.

Its 131K-token context window handles most long-document tasks — roughly 75,000 words or a full novel — without requiring chunking or retrieval augmentation. Median latency (P50 time-to-first-token) is approximately 25ms — fast enough for real-time interactive applications. Throughput is approximately 3200 tokens/second — fast enough to stream responses at reading speed with headroom for concurrent requests. Maximum output length is 8K tokens — adequate for most generation tasks including essays, code functions, and detailed responses.

llama-4-scout-17b-16e is best suited to speed, vision, cost. Supported capabilities include vision inputs (images alongside text), function calling / tool use for agentic workflows, streaming token output for real-time UIs. Vision support enables document analysis, screenshot understanding, diagram interpretation, and visual question answering — images are converted to tokens at the standard input rate based on resolution. Function calling support makes it suitable for building AI agents, structured data extraction pipelines, and any workflow requiring the model to invoke external APIs or return typed JSON. When evaluating this model for a specific workload, the most important factors are: context window (does your input fit?), output price (how long are your generations?), and latency (is this interactive or batch?). The pricing table above shows how llama-4-scout-17b-16e compares to similar models across these dimensions.

llama-4-scout-17b-16e sits in the efficient tier — optimized for speed and cost while maintaining solid quality, ideal for high-volume or latency-sensitive workloads. As an open-weights model, llama-4-scout-17b-16e can be self-hosted on GPU compute (see the GPU rental comparison on this site) to eliminate per-token API costs entirely — the break-even point versus API pricing depends on your volume and the GPU hourly rate. When comparing llama-4-scout-17b-16e to alternatives, use the similar models section on this page to find models at comparable price points with different capability tradeoffs. The intelligence score (68/100) reflects aggregate benchmark performance — a useful proxy for general capability, though real-world performance varies significantly by task type. Always test with your specific prompts and use case before committing to a model for production.