Mistral: Mistral Small 3

Mistral Small 3 is a text generation model from Mistral, released by the provider on 30 January 2025. It is a 24B-parameter language model optimized for low-latency performance across common AI tasks, and it is released under the Apache 2.0 license. Both pre-trained and instruction-tuned versions exist, and the instruction-tuned variant is the one served here under the model id mistralai/mistral-small-24b-instruct-2501.

The model works with text input and produces text output, so it fits standard chat and completion style workloads. It has a context window of 32768 tokens, which covers the prompt and the generated response together, and it can produce up to 16384 tokens in a single output. Its knowledge cutoff is 31 October 2023, so anything that happened after that date is outside what the model learned during training. The combination of a moderate parameter count and a 32k context makes it a general-purpose instruct model rather than a specialist in any single narrow task.

Because it is instruction-tuned and optimized for low latency, the model suits everyday tasks where response speed matters: chat assistants, summarization, rewriting, classification, extraction, and question answering over material that fits within the context window. The 16384-token output ceiling allows longer generations such as multi-paragraph drafts or structured documents, while the 32768-token context supports prompts that include a fair amount of reference text, for example a few documents or a long conversation history. Developers who need a dependable text model for high-volume, latency-sensitive endpoints are the intended audience.

You call it through api14's OpenAI-compatible API using the model id mistralai/mistral-small-24b-instruct-2501, which means existing OpenAI client libraries and request shapes work with only the model name changed. The model supports frequency_penalty, logit_bias, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_k, and top_p. Structured outputs and response_format are available for cases where you need the reply in a defined shape, and seed can help make sampling more reproducible. Sampling behaviour is controlled through temperature, top_p, top_k, and min_p, while repetition_penalty, frequency_penalty, and presence_penalty shape how much the model repeats itself.

Two practical constraints are worth noting before you commit. The knowledge cutoff of 31 October 2023 means the model has no built-in awareness of later events, so applications that depend on recent information need to supply it in the prompt. Input and output are text only, so if your workload involves images or audio, this model is not the right fit. And with a 32768-token context window, very large inputs will need chunking or retrieval rather than being passed in whole.