Particle Inference API

from openai import OpenAI

client = OpenAI(
    base_url="https://api.particle.ai/v1",
    api_key="pai-...",  # from platform.particle.ai
)

response = client.chat.completions.create(
    model="particle-1",
    messages=[{"role": "user", "content": "Explain KV caching in two sentences."}],
)
print(response.choices[0].message.content)

​

What’s available

​

Models

​

Core behavior

• Reasoning by default. All three models are reasoning models. If you omit reasoning_effort, requests run at the default effort (high) and reasoning tokens appear in usage.completion_tokens_details.reasoning_tokens. On DeepSeek and Particle-1, set "reasoning_effort": "none" to disable thinking. GLM always reasons — reasoning-off selectors are rejected with 400. See Reasoning controls.
• Safe sampling defaults. If you send no tail control (top_p, top_k, min_p), the gateway applies top_p=0.85 to keep long agentic generations stable. Any explicit value you send — including top_p=1.0 — is honored unchanged.
• Bounded output. If you omit both max_tokens and max_completion_tokens, the gateway supplies 32000 so an unbounded generation cannot run forever. Your explicit limits are preserved.
• Streaming with keep-alives. During long reasoning pauses, the stream carries real empty chat.completion.chunk frames after 15s of silence so client watchdogs and proxies don’t kill the connection. See Streaming.
• Prompt-cache affinity. Send prompt_cache_key to keep a conversation’s repeated prefix on the same server — large savings on multi-turn agentic workloads. See Prompt caching.
• Regular or Flex service. Set "service_tier": "flex" for 50% lower input, cached input, and output token prices with lower priority when busy. Use regular service for latency-sensitive requests, or Flex when you can tolerate waiting or retrying. See Service tiers.

​

Capacity and errors

​

Data handling