Phi-3.5 Mini Instruct
A 3.8B-parameter Phi-3.5 chat model from Microsoft, post-trained for multi-turn and multilingual conversation across 23 listed languages. Released under the MIT license.
Overview
Microsoft released Phi-3.5 Mini Instruct in August 2024 as an update to the June 2024 Phi-3 Mini. It is a 3.8B-parameter dense decoder-only Transformer on the Phi-3 Mini tokenizer, trained on 3.4T tokens with public data up to October 2023. Microsoft says the added post-training data improved multi-turn conversation and output in languages other than English.
The card lists 23 supported languages, Arabic, Hebrew, Korean and Turkish among them, and names commercial and research use as the intent. On multilingual MMLU, Microsoft reports 55.4, up from 51.08 for the June Phi-3 Mini 128K.
Choosing a variant: this is the smaller of the two Phi models on Wiro and the one to pick when replies must come in a language other than English. Phi-4 has 14B parameters and an English focus; use it for harder math and reasoning.
What you can build
- Support replies in a listed language such as Polish or Thai.
- Answer boxes that search first and let the model write from the results, a pairing Microsoft suggests because a 3.8B model has little room to store facts.
- Homework helpers for grade-school word problems. Microsoft reports 86.2 on GSM8K.
Inputs
A prompt is required. An optional system prompt can fix the reply language or format. To continue a conversation, send the same user_id and session_id each time; leave them empty and every call is independent.
Temperature defaults to 0.7 (allowed 0 to 5) and top_p to 0.95, with seed 123456. Sampling can be switched off with do_sample, and you can also adjust top_k, penalties for repetition and length, reply length (max_tokens, min_tokens) and stop sequences. Quantization is enabled by default and cuts memory use with reduced-precision weights.
Outputs
Generated text. The sample lists six Tower of London events with a story hook for each, and several are wrong: it puts Princess Elizabeth's imprisonment in 1514 and links it to Prince Arthur.
Recommended settings
- For math or extraction, copy Microsoft's usage example: do_sample off and temperature 0.
- Start a new session_id when a chat gets long; the card warns that very long sessions can turn repetitive or inconsistent.
- Ask the question up front in long prompts, because text over the input limit is removed from the end of your message.
- With max_tokens left at 0, a prompt of 2,048 tokens or less has room for a 2,048-token reply. Raise it for longer answers.
Limitations
- The checkpoint supports a 128K (131,072-token) window, but the input limit on Wiro is lower and covers the system prompt and chat history along with your new message. Over it, the oldest history messages go first, then the end of your latest message.
- Expect wrong dates and names unless you supply source text.
- Non-English output is weaker than English and uneven from one language to the next. Microsoft also flags safety gaps outside English.
Safety & compliance
The license is MIT. Microsoft asks developers to test the model for their own use case and language, and to follow applicable privacy and trade laws. Its red-teaming found the Phi-3.5 models may be more susceptible to long multi-turn jailbreaks in English and other languages, so public-facing chat needs its own safeguards.
API quick start
Run Phi-3.5-mini-instruct with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/microsoft/phi-3-5-mini-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



