Particle AI
Particle provides an OpenAI-compatible inference API for open-weight models, including its post-trained Particle-1 model, DeepSeek V4.1 Flash, and GLM 5.3 Flash. The service emphasizes low cost and token efficiency, with Particle-1 using roughly 45% fewer tokens on average across tasks. It offers a 1M-token context window and zero data retention for standard inference requests.
- Artificial Intelligence
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Developers building AI-powered applications face high inference costs and latency, limiting how often and where they can deploy capable models. The infrastructure required to run frontier-quality models is expensive and concentrated, while the quality gap between open-weight and closed models narrows, making efficient inference a key competitive factor.
Solution
Particle offers a fast, low-cost, and reliable AI inference layer for open-weight models through a single OpenAI-compatible API endpoint. The platform serves Particle-1, a model post-trained on DeepSeek V4.1 Flash to use significantly fewer tokens while maintaining accuracy, alongside DeepSeek V4.1 Flash and GLM 5.3 Flash. Developers can integrate via standard SDKs, with support for chat completions, streaming, tool use, structured outputs, and file references. For agentic use cases, a WebSocket-based Responses API enables mid-turn steering, allowing user updates during generation. The service operates under a zero data retention policy for standard inference, ensuring raw prompts and outputs are not stored.
Target Audience
Primary customers are software developers and AI teams building applications, agents, and tools that require high-volume, cost-efficient inference from open-weight models, particularly those focused on agentic coding, data engineering, and long-context reasoning tasks.
Features
- OpenAI-compatible API at `https://api.particle.ai/v1` supporting Chat Completions, Completions, and List Models endpoints
- 1M-token context window across all available models, enabling long-document and complex agentic workflows
- Particle-1 post-trained via reinforcement learning to reduce output tokens by 28.4%–52.2% and cost per task by 34.4%–51.6% compared to DeepSeek V4.1 Flash on software engineering benchmarks
- Responses WebSocket API with mid-turn steering for real-time user updates during generation
- Reasoning controls allowing developers to set `reasoning_effort` (including disabling thinking on Particle-1 and DeepSeek) and structured output support
- Zero data retention for standard inference, with no use of customer content for model training unless explicitly opted in