Description
GLiDE 27B is a sophisticated 27 billion parameter decision model developed by Fastino, designed to provide probabilistic answers to typed questions based on diverse input states. The model excels in scenarios requiring precise classification, routing, and scoring, rather than text generation. It processes input states that can include text, JSON, and images, offering a multimodal capability for comprehensive decision analysis.
Read More
Users can interact with GLiDE 27B through Fastino's hosted API, which supports decision requests via a POST endpoint. The API allows for various question types, including 'choice' (selecting one from up to 255 named options), 'noul' (yes/no questions returning P(yes)), and 'score' (assigning ordered levels). For each question, GLiDE 27B returns probabilities for every possible option, derived from a 255-way decision readout, ensuring no generative text is produced.
The model is engineered for efficiency, performing one forward pass per question and scoring all questions within a request, as well as concurrent requests, together in micro-batches. It supports long contexts, accommodating up to 40,000 tokens per question prompt by default, which includes state, question, options, and image tokens. Interfaces are available via a Python API and an HTTP server that communicates using the System One format.
GLiDE 27B operates with bf16 weights, requiring a GPU with 80 GB or more memory, such as the NVIDIA RTX PRO 6000 Blackwell Server Edition. It has demonstrated a median latency of 112 ms per request on this hardware. The model's architecture includes a 51.2 GB language model, a 0.92 GB vision encoder, and a 2.6 MB readout. Fastino provides an API key system and documentation for easy integration and deployment. The model is licensed under the Fastino Community License, with third-party components retaining their original licenses.
- CategoryAI Models & LLMs
- Pricing ModelContact sales
- DeploymentCloud
- PlatformsWeb
GLiDE 27B Highlights
27 Billion Parameters
Image-Text-to-Text Decision Model
Typed Decision Questions (Choice, Noul, Score)
Probabilistic Outputs for Every Option
No Text Generation (Non-Generative)
Multimodal Input (Text, JSON, Images)
Long Context Support (Up to 40,000 Tokens)
Python API and HTTP Server Interfaces
Micro-Batch Processing for Efficiency
bf16 Weights and Compute
Requires 80 GB+ GPU Memory
Fastino Hosted API Available
API Key Authentication
Supports Inline Images (Data URLs, Base64)
Configurable Batching and Concurrency
Getting Started with GLiDE 27B
Get API Key: Obtain an API key from Fastino to access the hosted GLiDE API.
Prepare Request: Format your decision request with a state (text, JSON, or multimodal with images) and one or more typed questions.
Send Request: Use the Fastino API endpoint (POST https://api.fastino.ai/v1/systemone) with the model `fastino/GLiDE`.
Receive Response: Get a response containing the most probable choice, noul probability, or score, along with probabilities for all options.
Integrate Locally (Optional): Follow repository instructions to run GLiDE 27B locally using Python or an HTTP server.
GLiDE 27B's Use Cases
- Classification Tasks
- Decision Routing
- Scoring Systems
- Multimodal Analysis
- Sentiment Analysis
- Content Moderation
- Customer Support Triage
- Survey Analysis
