picoLLM Inference Engine Python Binding

Made in Vancouver, Canada by Picovoice

picoLLM Inference Engine

picoLLM Inference Engine is a highly accurate and cross-platform SDK optimized for running compressed large language models. picoLLM Inference Engine is:

• Accurate; picoLLM Compression improves GPTQ by significant margins
• Private; LLM inference runs 100% locally.
• Cross-Platform
• Runs on CPU and GPU
• Free for open-weight models

Compatibility

• Python 3.9+
• Runs on Linux (x86_64), macOS (arm64, x86_64), Windows (x86_64, arm64), and Raspberry Pi (3, 4, 5).

Installation

pip3 install picollm

Models

picoLLM Inference Engine supports the following open-weight models. The models are on Picovoice Console.

• DeepSeek-OCR-2
  • deepseek-ocr-2
• EmbeddingGemma
  • embeddinggemma-300m
• Gemma
  • gemma-2b
  • gemma-2b-it
  • gemma-7b
  • gemma-7b-it
• Gemma3
  • gemma-3-270m
  • gemma-3-270m-it
• Llama-2
  • llama-2-7b
  • llama-2-7b-chat
  • llama-2-13b
  • llama-2-13b-chat
  • llama-2-70b
  • llama-2-70b-chat
• Llama-3
  • llama-3-8b
  • llama-3-8b-instruct
  • llama-3-70b
  • llama-3-70b-instruct
• Llama-3.2
  • llama3.2-1b-instruct
  • llama3.2-3b-instruct
• Mistral
  • mistral-7b-v0.1
  • mistral-7b-instruct-v0.1
  • mistral-7b-instruct-v0.2
• Mixtral
  • mixtral-8x7b-v0.1
  • mixtral-8x7b-instruct-v0.1
• Phi-2
  • phi2
• Phi-3
  • phi3
• Phi-3.5
  • phi3.5
• Qwen3-VL
  • qwen3-vl-2b-it

AccessKey

AccessKey is your authentication and authorization token for deploying Picovoice SDKs, including picoLLM. Anyone who is using Picovoice needs to have a valid AccessKey. You must keep your AccessKey secret. You would need internet connectivity to validate your AccessKey with Picovoice license servers even though the LLM inference is running 100% offline and completely free for open-weight models. Everyone who signs up for Picovoice Console receives a unique AccessKey.

Usage

Text models

Create an instance of the engine and generate a prompt completion:

import picollm

pllm = picollm.create(
    access_key='${ACCESS_KEY}',
    model_path='${MODEL_PATH}')

res = pllm.generate(prompt='${PROMPT}')
print(res.completion)

Replace ${ACCESS_KEY} with yours obtained from Picovoice Console, ${MODEL_PATH} with the path to a model file downloaded from Picovoice Console, and ${PROMPT} with a prompt string.

Instruction-tuned models (e.g., llama-3-8b-instruct, llama-2-7b-chat, and gemma-2b-it) have a specific chat template. You can either directly format the prompt or use a dialog helper:

dialog = pllm.get_dialog()
dialog.add_human_request(prompt)

res = pllm.generate(prompt=dialog.prompt())
dialog.add_llm_response(res.completion)
print(res.completion)

To interrupt completion generation before it has finished:

pllm.interrupt()

Finally, when done, be sure to release the resources explicitly:

pllm.release()

Vision models

To run a VLM such as qwen3-vl-2b-it:

res = pllm.generate_with_image(
    prompt='${PROMPT}',
    image_width=${IMAGE_NUM_PIXELS_WIDTH},
    image_height=${IMAGE_NUM_PIXELS_HEIGHT},
    image=${IMAGE_DATA});
print(res.completion)

Replace ${PROMPT} with a text prompt. For the image, you will need to get image height and width in number of pixels and the raw pixel values of the image in 8-bit, RGB format.

OCR models

To run an OCR model such as deepseek-ocr-2:

res = pllm.generate_ocr(
    image_width=${IMAGE_NUM_PIXELS_WIDTH},
    image_height=${IMAGE_NUM_PIXELS_HEIGHT},
    image=${IMAGE_DATA});
print(res.completion)

For the image, you will need to get image height and width in number of pixels and the raw pixel values of the image in 8-bit, RGB format.

Embedding models

To run an embedding model such as embeddinggemma-300m:

res = pllm.generate_embeddings(prompt='${PROMPT}');
for embedding in range(len(res)):
  print(embedding)

Replace ${PROMPT} with a text prompt that you want to generate embeddings for.

Demos

picollmdemo provides command-line utilities for LLM completion and chat using picoLLM.