unsloth

/

Mistral-Small-3.2-24B-Instruct-2506-GGUF

like 35

Follow

Unsloth AI 6.56k

So basically what they talk about here: https://github.com/lmstudio-ai/lmstudio-bug-tracker/issues/710 you have already implemented?

Thank you for putting the effort in, it is really appreciated!

So basically what they talk about here: https://github.com/lmstudio-ai/lmstudio-bug-tracker/issues/710 you have already implemented?

Thank you for putting the effort in, it is really appreciated!

Yes correct

@simonw You might be interested in these quants where we managed to add tool calling support and also use the exact system prompt as suggested by Mistral!

In llama.cpp it appears to be failing tool calls, not sure if im doing something wrong here

"Failed to infer a tool call example (possible template bug)"

Model: unsloth_Mistral-Small-3.2-24B-Instruct-2506-GGUF_Mistral-Small-3.2-24B-Instruct-2506-UD-Q8_K_XL.gguf

When used for tool calling, I always see this output from Mistral-Small-3.2-24B-Instruct-2506-UD-Q6_K_XL.gguf.
I am trying this with Jan Beta and aichat.

In Tekken 7, Mistral is a powerful character known for her agility and versatility. She is a French fighter who uses a mix of martial arts, including Muay Thai and Capoeira, to overwhelm her opponents. Mistral's moveset includes a variety of kicks and spins, making her a formidable opponent in both offense and defense. Her signature moves, such as the 'Mistral Kick' and 'French Kiss,' are particularly effective in close combat. Additionally, her ability to quickly change directions and evade attacks makes her a challenging character to counter. Overall, Mistral is a dynamic and exciting character to play in Tekken 7, offering a unique blend of speed, power, and technique.

@qikchen Interesting - you're right - interestingly normal hugging face inference works fine for "What's the temperature in San Francisco now? How about tomorrow?". We get "[TOOL_CALLS]get_current_temperature[ARGS]{"location": "San Francisco, California, USA", "unit": "celsius"}[TOOL_CALLS]get_temperature_date[ARGS]{"location": "San Francisco, California, USA", "date": "2025-06-24", "unit": "celsius"}</s>"

#  pip install unsloth
from unsloth import FastLanguageModel
import torch
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Mistral-Small-3.2-24B-Instruct-2506",
    max_seq_length = 2048,
    load_in_4bit = True,
)
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_current_temperature",
            "description": "Get current temperature at a location.",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": 'The location to get the temperature for, in the format "City, State, Country".',
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": 'The unit to return the temperature in. Defaults to "celsius".',
                    },
                },
                "required": ["location"],
            },
        },
    },
    {
        "type": "function",
        "function": {
            "name": "get_temperature_date",
            "description": "Get temperature at a location and date.",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": 'The location to get the temperature for, in the format "City, State, Country".',
                    },
                    "date": {
                        "type": "string",
                        "description": 'The date to get the temperature for, in the format "Year-Month-Day".',
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": 'The unit to return the temperature in. Defaults to "celsius".',
                    },
                },
                "required": ["location", "date"],
            },
        },
    },
]
messages = [
    {"role": "user", "content": "What's the temperature in San Francisco now? How about tomorrow?"},
]

inputs = tokenizer.tokenizer.apply_chat_template(messages, tools = tools, return_tensors = "pt").to("cuda")
from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer, skip_prompt = True)
_ = model.generate(input_ids = inputs, streamer = text_streamer, max_new_tokens = 128, temperature = 0, do_sample = False)

For llama.cpp, doing the below works as well, we get "get_current_temperature{"location": "San Francisco, California, USA"}"

./llama.cpp/llama-cli     --model Mistral-Small-3.2-24B-Instruct-2506-GGUF/Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf     -ngl 99     --temp 0.0     --device CUDA0     -no-cnv     --prompt "[SYSTEM_PROMPT]You are Mistral-Small-3.2-24B-Instruct-2506, a Large Language Model (LLM) created by Mistral AI, a French startup headquartered in Paris.\nYou power an AI assistant called Le Chat.\nYour knowledge base was last updated on 2023-10-01.\nThe current date is 2025-06-23.\n\nWhen you\'re not sure about some information or when the user\'s request requires up-to-date or specific data, you must use the available tools to fetch the information. Do not hesitate to use tools whenever they can provide a more accurate or complete response. If no relevant tools are available, then clearly state that you don\'t have the information and avoid making up anything.\nIf the user\'s question is not clear, ambiguous, or does not provide enough context for you to accurately answer the question, you do not try to answer it right away and you rather ask the user to clarify their request (e.g. \"What are some good restaurants around me?\" => \"Where are you?\" or \"When is the next flight to Tokyo\" => \"Where do you travel from?\").\nYou are always very attentive to dates, in particular you try to resolve dates (e.g. \"yesterday\" is 2025-06-22) and when asked about information at specific dates, you discard information that is at another date.\nYou follow these instructions in all languages, and always respond to the user in the language they use or request.\nNext sections describe the capabilities that you have.\n\n# WEB BROWSING INSTRUCTIONS\n\nYou cannot perform any web search or access internet to open URLs, links etc. If it seems like the user is expecting you to do so, you clarify the situation and ask the user to copy paste the text directly in the chat.\n\n# MULTI-MODAL INSTRUCTIONS\n\nYou have the ability to read images, but you cannot generate images. You also cannot transcribe audio files or videos.\nYou cannot read nor transcribe audio files or videos.\n\n# TOOL CALLING INSTRUCTIONS\n\nYou may have access to tools that you can use to fetch information or perform actions. You must use these tools in the following situations:\n\n1. When the request requires up-to-date information.\n2. When the request requires specific data that you do not have in your knowledge base.\n3. When the request involves actions that you cannot perform without tools.\n\nAlways prioritize using tools to provide the most accurate and helpful response. If tools are not available, inform the user that you cannot perform the requested action at the moment.[/SYSTEM_PROMPT][AVAILABLE_TOOLS][{\"type\": \"function\", \"function\": {\"name\": \"get_current_temperature\", \"description\": \"Get current temperature at a location.\", \"parameters\": {\"type\": \"object\", \"properties\": {\"location\": {\"type\": \"string\", \"description\": \"The location to get the temperature for, in the format \\\"City, State, Country\\\".\"}, \"unit\": {\"type\": \"string\", \"enum\": [\"celsius\", \"fahrenheit\"], \"description\": \"The unit to return the temperature in. Defaults to \\\"celsius\\\".\"}}, \"required\": [\"location\"]}}}, {\"type\": \"function\", \"function\": {\"name\": \"get_temperature_date\", \"description\": \"Get temperature at a location and date.\", \"parameters\": {\"type\": \"object\", \"properties\": {\"location\": {\"type\": \"string\", \"description\": \"The location to get the temperature for, in the format \\\"City, State, Country\\\".\"}, \"date\": {\"type\": \"string\", \"description\": \"The date to get the temperature for, in the format \\\"Year-Month-Day\\\".\"}, \"unit\": {\"type\": \"string\", \"enum\": [\"celsius\", \"fahrenheit\"], \"description\": \"The unit to return the temperature in. Defaults to \\\"celsius\\\".\"}}, \"required\": [\"location\", \"date\"]}}}][/AVAILABLE_TOOLS][INST]What\'s the temperature in San Francisco now? How about tomorrow?[/INST]"

Using --verbose, we see the tokens are produced correctly:

[ '[TOOL_CALLS]':9 ]
[ 'get':1689 ]
 [ '_current':45971 ]    
 [ '_t':2480 ]
 [ 'emperature':32294 ]                                                                 
 [ '[ARGS]':32 ]
[ '{"':19227 ]         
[ 'location':17611 ]                            
 [ '":':2811 ]                    
 [ ' "':1429 ]                    
[ 'San':24619 ]                 
[ ' Fran

Yes, I see the correct tool call happening with llama-cli on UD-Q6_K_XL as well. Perhaps the system prompts from aichat and Jan Beta aren't working?
I guess I can try llm by Simon Willison later.

When used for tool calling, I always see this output from Mistral-Small-3.2-24B-Instruct-2506-UD-Q6_K_XL.gguf.
I am trying this with Jan Beta and aichat.

In Tekken 7, Mistral is a powerful character known for her agility and versatility. She is a French fighter who uses a mix of martial arts, including Muay Thai and Capoeira, to overwhelm her opponents. Mistral's moveset includes a variety of kicks and spins, making her a formidable opponent in both offense and defense. Her signature moves, such as the 'Mistral Kick' and 'French Kiss,' are particularly effective in close combat. Additionally, her ability to quickly change directions and evade attacks makes her a challenging character to counter. Overall, Mistral is a dynamic and exciting character to play in Tekken 7, offering a unique blend of speed, power, and technique.

This is the same as what Im seeing as well with llama.cpp and owui

The Mistral AI 7 large language model (LLM) is a powerful tool for natural language processing tasks. It is designed to understand and generate human-like text based on the input it receives. The model is trained on a diverse range of internet text, which allows it to have a broad knowledge base and the ability to

@qikchen @mumblerit I found a solution! The culprit was llama.cpp was adding 1970 to the date, in which I did not consider - so I added a fallback - I confirmed in llama.cpp it should work as expected now! :) Please retry them if you guys have time! Hope it works!

@Dampfinchen Also best to redownload them to get fixed tool calling for llama.cpp!

I assume this was your change?

    {%- elif yesterday_year == '1970' %}
        {#- Stop llama_cpp from erroring out #}
        {%- set yesterday_day = '29' %}
    {%- else %}
        {{- raise_exception('Unsloth custom template does not support years > 2032. Error year = [' + yesterday_year + ']') }}

Im still seeing the same output when attempting a tool call

In Tekken 7, Mistral is a character who is a powerful and agile fighter. She is known for her speed and precision, making her a formidable opponent in the game. Mistral's moveset includes a variety of kicks and punches, as well as some unique special moves that can catch opponents off guard. Her playstyle is often described as aggressive and unpredictable, as she can quickly switch between offense and defense. Mistral's design is inspired by French culture, and she often wears a beret and a striped shirt in her outfits. Overall, her character is a popular choice among Tekken players due to her dynamic fighting style and stylish appearance.

@danielhanchen Can you share just the template? Would like to try it out before a new large download.

@qikchen Oh click on chat template on the model page and copy paste it into a new file called chat_template.jinja, then call llama-server --chat-template-file chat_template.jinja --jinja

@mumblerit Oh there's 2 changes so if it's just that one there's one more

@danielhanchen Thanks! It worked.

confirm specifying using the chat template manually fixes it