NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI · #1805 most downloaded on PyPI
Mistral-common is a library of common utilities for Mistral AI.
Last release 12 days ago
22 Sep 2026
Ships fairly regularly
a new release about every 4 weeks
Most releases are documented
notes for 40 of 50 stable releases
Nothing withdrawn
no release was ever pulled
2 years old
50 releases · first in 2024
Deprecate Tokenized.text in favor of tokenizer.decode by @juliendenize in #282
Full Changelog: v1.11.7...v1.12.0
One column per month.
Update AGENTS.md structure and add explicit keyword args rule by @juliendenize in https://github.com/mistralai/mistral-common/pull/270
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.6...v1.11.7
Full Changelog: v1.11.6...v1.11.7
Restore deprecated reasoning parameter in select_jinja_template by @juliendenize in https://github.com/mistralai/mistral-common/pull/266
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.5...v1.11.6
Full Changelog: v1.11.5...v1.11.6
Fix multi-image content ordering by @juliendenize in https://github.com/mistralai/mistral-common/pull/254
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.4...v1.11.5
Full Changelog: v1.11.4...v1.11.5
Deprecate RawAudio in favor of str | bytes by @juliendenize in https://github.com/mistralai/mistral-common/pull/227
AudioChunk.to_openai() serialization for raw audio bytes and prefixed base64 audio strings. by @haoruilee in https://github.com/mistralai/mistral-common/pull/245Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.3...v1.11.4
AudioChunk.to_openai() serialization for raw audio bytes and prefixed base64 audio strings. by @haoruilee in #245Full Changelog: v1.11.3...v1.11.4
Raise multiple format of reasoning for from_openai by @juliendenize in https://github.com/mistralai/mistral-common/pull/224
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.2...v1.11.3
Full Changelog: v1.11.2...v1.11.3
Add test and docstring to get_validator by @juliendenize in https://github.com/mistralai/mistral-common/pull/219
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.1...v1.11.2
Full Changelog: v1.11.1...v1.11.2
This Patch allows usage of user message after tool message. It also makes from_openai less strict to make mistral-common integrations in other framewo
This Patch allows usage of user message after tool message. It also makes from_openai less strict to make mistral-common integrations in other frameworks smoother.
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.11.0...v1.11.1
This Patch allows usage of user message after tool message. It also makes from_openai less strict to make mistral-common integrations in other frameworks smoother.
Full Changelog: v1.11.0...v1.11.1
Make use of lark grammar to guide your model in generating valid reasoning traces with or without tool calls !
Mistral Guidance is out !
Make use of lark grammar to guide your model in generating valid reasoning traces with or without tool calls !
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.10.0...v1.11.0
Mistral Guidance is out !
Make use of lark grammar to guide your model in generating valid reasoning traces with or without tool calls !
Full Changelog: v1.10.0...v1.11.0
Allow System Prompt with Audio for v13 by @juliendenize in https://github.com/mistralai/mistral-common/pull/184
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.9.1...v1.10.0
Full Changelog: v1.9.1...v1.10.0
Refactor online streaming processing and allow for dynamic streaming delay
Refactor online streaming processing and allow for dynamic streaming delay
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.9.0...v1.9.1
Refactor online streaming processing and allow for dynamic streaming delay
Full Changelog: v1.9.0...v1.9.1
from mistral_common.audio import Audio from mistral_common.protocol.instruct.chunk import RawAudio from mistral_common.protocol.transcription.request
import numpy as np
from mistral_common.audio import Audio
from mistral_common.protocol.instruct.chunk import RawAudio
from mistral_common.protocol.transcription.request import (
StreamingMode,
TranscriptionRequest,
)
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
# 1. Load the tokenizer with audio support
tokenizer = MistralTokenizer.from_hf_hub(
"mistralai/Voxtral-Mini-4B-Realtime-2602"
)
# 2. Create sample audio data (or load from a file)
sampling_rate = 16_000
duration_s = 2.0
audio_array = np.random.uniform(-1, 1, size=int(duration_s * sampling_rate)).astype(np.float32)
audio = Audio(
audio_array=audio_array,
sampling_rate=sampling_rate,
format="wav",
)
# 3. Create the streaming transcription request
request = TranscriptionRequest(
audio=RawAudio(
data=audio.to_base64("wav"),
format="wav",
),
streaming=StreamingMode.ONLINE, # or StreamingMode.OFFLINE
language=None,
)
# 4. Encode the request
tokenized = tokenizer.encode_transcription(request)
# 5. Access the results
print(f"Tokens: {tokenized.tokens}")
print(f"Number of tokens: {len(tokenized.tokens)}")
print(f"Number of audio segments: {len(tokenized.audios)}")
See https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602 for more info.
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.7...v1.9.0
Add new token logic asrstr by @patrickvonplaten in https://github.com/mistralai/mistral-common/pull/172
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.7...v1.8.8
Remove the index field from assistant tool_calls. by @tobrun in https://github.com/mistralai/mistral-common/pull/165
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.6...v1.8.7
Remove deprecated imports in docs. by @juliendenize in https://github.com/mistralai/mistral-common/pull/138
revision and token to hf_api by @juliendenize in https://github.com/mistralai/mistral-common/pull/149Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.5...v1.8.6
Make model field optional in TranscriptionRequest by @juliendenize in https://github.com/mistralai/mistral-common/pull/128
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.4...v1.8.5
Update experimental.md by @juliendenize in https://github.com/mistralai/mistral-common/pull/124
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.3...v1.8.4
Add a FastAPI app by @juliendenize in https://github.com/mistralai/mistral-common/pull/113
We released an experimental REST API leveraging Fast API to handle requests from tokenization, through generation via calls to an engine, to detokenization.
For a detailed documentation see [https://mistralai.github.io/mistral-common/usage/experimental/].
Here is how to launch the server:
pip install mistral-common[server]
mistral_common serve mistralai/Magistral-Small-2507 \
--host 127.0.0.1 --port 8000 \
--engine-url http://127.0.0.1:8080 --engine-backend llama_cpp \
--timeout 60
Then you can see the Swagger at: http://localhost:8000.
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.2...v1.8.3
Add think chunk by @juliendenize in https://github.com/mistralai/mistral-common/pull/122
Now you can use TextChunk and ThinkChunk in your SystemMessage or AssistantMessage:
from mistral_common.protocol.instruct.messages import SystemMessage, TextChunk, ThinkChunk
system_message = SystemMessage(
content = [
TextChunk(text="First draft your thinking process (inner monologue) until you arrive at a response. Format your response using Markdown, and use LaTeX for any mathematical equations. Write both your thoughts and the response in the same language as the input.\n\nYour thinking process must follow the template below:"),
ThinkChunk(
thinking="Your thoughts or/and draft, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate the response. Use the same language as the input.",
closed=True,
),
TextChunk(text="Here, provide a self-contained response.")
],
)
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.1...v1.8.2
Add AudioURLChunk by @juliendenize in https://github.com/mistralai/mistral-common/pull/120
Now you can use http(s) URLs, file paths and base64 string (without specifying format) in your content chunks thanks to AudioURLChunk !
from mistral_common.protocol.instruct.messages import AudioURL, AudioURLChunk, TextChunk, UserMessage
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
repo_id = "mistralai/Voxtral-Mini-3B-2507"
tokenizer = MistralTokenizer.from_hf_hub(repo_id)
text_chunk = TextChunk(
text="Wat do you think about this audio?"
)
user_msg = UserMessage(
content=[
AudioURLChunk(audio_url=AudioURL(url="https://freewavesamples.com/files/Ouch-6.wav")),
text_chunk,
]
)
request = ChatCompletionRequest(messages=[user_msg])
tokenized = tokenizer.encode_chat_completion(request)
# pass tokenized.tokens to your favorite audio model
print(tokenized.tokens)
print(tokenized.audios)
# print text to visually see tokens
print(tokenized.text)
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.8.0...v1.8.1
[Audio] Add audio by @patrickvonplaten in https://github.com/mistralai/mistral-common/pull/119
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.7.0...v1.8.0
from mistral_common.protocol.instruct.messages import TextChunk, AudioChunk, UserMessage, AssistantMessage, RawAudio
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.audio import Audio
from huggingface_hub import hf_hub_download
repo_id = "mistralai/voxtral-mini"
tokenizer = MistralTokenizer.from_hf_hub(repo_id)
obama_file = hf_hub_download("patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset")
bcn_file = hf_hub_download("patrickvonplaten/audio_samples", "bcn_weather.mp3", repo_type="dataset")
def file_to_chunk(file: str) -> AudioChunk:
audio = Audio.from_file(file, strict=False)
return AudioChunk.from_audio(audio)
text_chunk = TextChunk(text="Which speaker do you prefer between the two? Why? How are they different from each other?")
user_msg = UserMessage(content=[file_to_chunk(obama_file), file_to_chunk(bcn_file), text_chunk]).to_openai()
request = ChatCompletionRequest(messages=[user_msg])
tokenized = tokenizer.encode_chat_completion(request)
# pass tokenized.tokens to your favorite audio model
print(tokenized.tokens)
print(tokenized.audios)
# print text to visually see tokens
print(tokenized.text)
from mistral_common.protocol.transcription.request import TranscriptionRequest
from mistral_common.protocol.instruct.messages import RawAudio
from mistral_common.audio import Audio
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from huggingface_hub import hf_hub_download
repo_id = "mistralai/voxtral-mini"
tokenizer = MistralTokenizer.from_hf_hub(repo_id)
obama_file = hf_hub_download("patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset")
audio = Audio.from_file(obama_file, strict=False)
audio = RawAudio.from_audio(audio)
request = TranscriptionRequest(model=repo_id, audio=audio, language="en")
tokenized = tokenizer.encode_transcription(request)
# pass tokenized.tokens to your favorite audio model
print(tokenized.tokens)
print(tokenized.audios)
# print text to visually see tokens
print(tokenized.text)
[Naming] Rename multi-modal to image by @patrickvonplaten in https://github.com/mistralai/mistral-common/pull/114
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.6.3...v1.7.0
Nothing published for this version
Improve decode and deprecate to_string by @juliendenize in https://github.com/mistralai/mistral-common/pull/99
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.6.0...v1.6.3
Ensure that pypi version includes tokenizer files.
Ensure that pypi version includes tokenizer files.
Nothing published for this version
Constantize model to tokenizer mapping (for external import) by @djsaunde in https://github.com/mistralai/mistral-common/pull/86
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.5.6...v1.6.0
[Tokenizer >= v7] Allow content and tool calls to be defined together by @patrickvonplaten in https://github.com/mistralai/mistral-common/pull/82
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.5.5...v1.5.6
Deprecate runtime special tokens for tekkenizer. by @juliendenize in https://github.com/mistralai/mistral-common/pull/79
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.5.4...v1.5.5
Add spatial_merge_size for https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503
Add spatial_merge_size for https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503
Nothing published for this version
Nothing published for this version
[From model] Make from model much stricter by @patrickvonplaten in https://github.com/mistralai/mistral-common/pull/65
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.5.0...v1.5.1
Mistral's newest tokenizer has two major improvements:
Mistral's newest tokenizer has two major improvements:
Similar to other tokenization schemes the system prompt is now treated as a "normal" message encapsulated by [SYSTEM_PROMPT] ...[\SYSTEM_PROMPT]
E.g.
from mistral_common.protocol.instruct.messages import (
UserMessage,
SystemMessage,
AssistantMessage,
)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
# Load Mistral tokenizer
tokenizer = MistralTokenizer.v7()
# Tokenize a list of messages
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
messages=[
SystemMessage(content="You are a funny AI assistant. Always make jokes."),
UserMessage(content="What's the weather like today in Paris"),
],
model="joker",
)
)
tokens, text = tokenized.tokens, tokenized.text
print(text)
# <s>[SYSTEM_PROMPT]▁You▁are▁a▁funny▁AI▁assistant.▁Always▁make▁jokes.[/SYSTEM_PROMPT][INST]▁What's▁the▁weather▁like▁today▁in▁Paris[/INST]
A new [TOOL_CONTENT] is added if trained with correctly should improve the accuracy of function calling.
from mistral_common.protocol.instruct.messages import (
UserMessage,
SystemMessage,
AssistantMessage,
ToolMessage
)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.protocol.instruct.tool_calls import (
Function,
Tool,
)
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
# Load Mistral tokenizer
tokenizer = MistralTokenizer.v7()
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
tools=[
Tool(
function=Function(
name="get_current_weather",
description="Get the current weather",
parameters={
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
"required": ["location", "format"],
},
)
)
],
messages=[
UserMessage(content="What's the weather like today in Paris"),
AssistantMessage(content="", tool_calls=[
{
"id": "bbc5b7ede",
"type": "function",
"function": {
"name": "weather",
"arguments": '{"location": "Paris", "format": "celsius"}',
},
}
]),
ToolMessage(content="24 degrees celsius", tool_call_id="bbc5b7ede"),
],
model="joker",
)
)
tokens, text = tokenized.tokens, tokenized.text
# Count the number of tokens
print(text)
# <s>[AVAILABLE_TOOLS]▁[{"type":▁"function",▁"function":▁{"name":▁"get_current_weather",▁"description":▁"Get▁the▁current▁weather",▁"parameters":▁{"type":▁"object",▁"properties":▁{"location":▁{"type":▁"string",▁"description":▁"The▁city▁and▁state,▁e.g.▁San▁Francisco,▁CA"},▁"format":▁{"type":▁"string",▁"enum":▁["celsius",▁"fahrenheit"],▁"description":▁"The▁temperature▁unit▁to▁use.▁Infer▁this▁from▁the▁users▁location."}},▁"required":▁["location",▁"format"]}}}][/AVAILABLE_TOOLS][INST]▁What\'s▁the▁weather▁like▁today▁in▁Paris[/INST][TOOL_CALLS]▁[{"name":▁"weather",▁"arguments":▁{"location":▁"Paris",▁"format":▁"celsius"},▁"id":▁"bbc5b7ede"}]</s>[TOOL_RESULTS]▁bbc5b7ede[TOOL_CONTENT]▁24▁degrees▁celsius[/TOOL_RESULTS]'
Make sure broken user envs of cv2 (which sadly happens more often than not) don't impede users from using text-only models.
Make sure broken user envs of cv2 (which sadly happens more often than not) don't impede users from using text-only models.
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.4.3...v1.4.4
As per discussion: https://github.com/vllm-project/vllm/issues/8650 make cv2 optional.
As per discussion: https://github.com/vllm-project/vllm/issues/8650 make cv2 optional.
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.4.2...v1.4.3
Make sure to send user agent for downloading pictures that require a user agent. E.g.:
Make sure to send user agent for downloading pictures that require a user agent. E.g.:
from mistral_common.protocol.instruct.messages import (
UserMessage,
TextChunk,
ImageURLChunk,
ImageChunk,
)
from PIL import Image
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
tokenizer = MistralTokenizer.from_model("pixtral")
url_dog = "https://picsum.photos/id/237/200/300"
url_mountain = "https://picsum.photos/seed/picsum/200/300"
url1 = "https://upload.wikimedia.org/wikipedia/commons/d/da/2015_Kaczka_krzy%C5%BCowka_w_wodzie_%28samiec%29.jpg"
url2 = "https://upload.wikimedia.org/wikipedia/commons/7/77/002_The_lion_king_Snyggve_in_the_Serengeti_National_Park_Photo_by_Giles_Laurent.jpg"
# tokenize image urls and text
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
messages=[
UserMessage(
content=[
TextChunk(text="Can this animal"),
ImageURLChunk(image_url=url1),
TextChunk(text="live here?"),
ImageURLChunk(image_url=url2),
]
)
],
model="pixtral",
)
)
tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images
# Count the number of tokens
print("# tokens", len(tokens))
print("# images", len(images))
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.4.1...v1.4.2
cv2 resize gives significantly better results when running pixtral in inference as compared to PIL hence we're making a patch release to resize images
cv2 resize gives significantly better results when running pixtral in inference as compared to PIL hence we're making a patch release to resize images using cv2 as shown here: https://github.com/mistralai/mistral-common/commit/bae45b221a1b9de00c59acb44e79eeeea5d844c6
Mistral common has image support! You can now pass images and URLs alongside text into the user message.
Mistral common has image support! You can now pass images and URLs alongside text into the user message.
pip install --upgrade mistral_common
You can encode images as follows
from mistral_common.protocol.instruct.messages import (
UserMessage,
TextChunk,
ImageURLChunk,
ImageChunk,
)
from PIL import Image
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
tokenizer = MistralTokenizer.from_model("pixtral")
image = Image.new('RGB', (64, 64))
# tokenize images and text
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
messages=[
UserMessage(
content=[
TextChunk(text="Describe this image"),
ImageChunk(image=image),
]
)
],
model="pixtral",
)
)
tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images
# Count the number of tokens
print("# tokens", len(tokens))
print("# images", len(images))
You can pass image url which will be automatically downloaded
url_dog = "https://picsum.photos/id/237/200/300"
url_mountain = "https://picsum.photos/seed/picsum/200/300"
# tokenize image urls and text
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
messages=[
UserMessage(
content=[
TextChunk(text="Can this animal"),
ImageURLChunk(image_url=url_dog),
TextChunk(text="live here?"),
ImageURLChunk(image_url=url_mountain),
]
)
],
model="pixtral",
)
)
tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images
# Count the number of tokens
print("# tokens", len(tokens))
print("# images", len(images))
You can also pass image encoded as base64
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
messages=[
UserMessage(
content=[
TextChunk(text="What is this?"),
ImageURLChunk(image_url="data:image/jpeg;base64,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"),
]
)
],
model="pixtral",
)
)
tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images
# Count the number of tokens
print("# tokens", len(tokens))
print("# images", len(images))
In this patch release the pydantic requirement is loosened to be <= 3.0.0
In this patch release the pydantic requirement is loosened to be <= 3.0.0
as noticed in multiple issues, e.g.:
Nothing published for this version
Nothing published for this version
The new Tekkenizer class is based on Open AI's tiktoken and supports the new Mistral-Nemo model.
Tekkenizer
The new Tekkenizer class is based on Open AI's tiktoken and supports the new Mistral-Nemo model.
Tekkenizer always makes use of version 3 or higher.
Examples:
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
tokenizer = MistralTokenizer.v3(is_tekken=True)
tokenizer = MistralTokenizer.from_model("...")
Function calling (just like before)
# Import needed packages:
from mistral_common.protocol.instruct.messages import (
UserMessage,
)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from mistral_common.protocol.instruct.tool_calls import (
Function,
Tool,
)
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
# Load Mistral tokenizer
model_name = "..."
tokenizer = MistralTokenizer.from_model(model_name)
# Tokenize a list of messages
tokenized = tokenizer.encode_chat_completion(
ChatCompletionRequest(
tools=[
Tool(
function=Function(
name="get_current_weather",
description="Get the current weather",
parameters={
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location.",
},
},
"required": ["location", "format"],
},
)
)
],
messages=[
UserMessage(content="What's the weather like today in Paris"),
],
model=model_name,
)
)
tokens, text = tokenized.tokens, tokenized.text
# Count the number of tokens
print(len(tokens))
Full Changelog: https://github.com/mistralai/mistral-common/compare/v1.3.0...v1.3.1
Nothing published for this version
As noticed here: https://huggingface.co/mistralai/Codestral-22B-v0.1/discussions/10
As noticed here: https://huggingface.co/mistralai/Codestral-22B-v0.1/discussions/10
The wrong tokenizer was used for FIM. This patch release fixes that so that the following works correctly:
from mistral_common.tokens.tokenizers.base import FIMRequest
from mistral_common_private.tokens.tokenizers.mistral import MistralTokenizer
tokenizer = MistralTokenizer.v3()
tokenized = tokenizer.encode_fim(FIMRequest(prompt="def f(", suffix="return a + b"))
assert tokenized.text == "<s>[SUFFIX]return▁a▁+▁b[PREFIX]▁def▁f("
Fill-in-the-middle (FIM) with [SUFFIX] and [PREFIX] logic is added to allow building code completion workflows such as:
Fill-in-the-middle (FIM) with [SUFFIX] and [PREFIX] logic is added to allow building code completion workflows such as:
from mistral_inference.model import Transformer
from mistral_inference.generate import generate
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.tokens.instruct.request import FIMRequest
tokenizer = MistralTokenizer.v3()
model = Transformer.from_folder("~/codestral-22B-240529")
prefix = """def add("""
suffix = """ return sum"""
request = FIMRequest(prompt=prefix, suffix=suffix)
tokens = tokenizer.encode_fim(request).tokens
out_tokens, _ = generate([tokens], model, max_tokens=256, temperature=0.0, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
result = tokenizer.decode(out_tokens[0])
middle = result.split(suffix)[0].strip()
print(middle)
Adds improved function calling validator
Nothing published for this version
Nothing published for this version
Nothing published for this version
Your coding agent can read these notes before it upgrades. Set up the MCP server →