Throughput (tokens / sec)
avg. 314tokens/s
Low-latency reasoning on a tiny budget
Run instantly. Pay only for what you use.
GPT OSS 20B is OpenAI's compact open-weight reasoning model, released under Apache 2.0: a 21B-parameter MoE with 3.6B active per token and a 131K-token context window, trained on the Harmony response format with configurable reasoning effort — low, medium, or high.
It is the price-performance floor for real reasoning: at high effort it reaches 98.7% on AIME 2025 (with tools) and 71.5% on GPQA — OpenAI's published evals put it near o3-mini — from a model small enough to run on 16GB devices.
That makes structured work cheap: Harmony-trained function calling and strict JSON output with top-15 output speed among the open models measured by Artificial Analysis.
3.6Bactive of 21B total
MoE routing delivers real reasoning at the smallest serving footprint.
3reasoning effort levels
Low, medium, and high — accuracy scales smoothly as effort rises.
131Ktoken context window
Long documents and full tool-call transcripts fit in a single request.
GPT OSS 20B is strongest where volume and latency dominate — classification, routing, extraction, and assistant backends that need reasoning quality without reasoning prices.
A one-line system-message setting trades latency for depth — smooth accuracy scaling across low, medium, and high.
98.7% on AIME 2025 with tools at high effort — from a model that fits on 16GB devices.
Harmony-trained function calling and strict JSON — tool arguments that parse the first time, at top-15 open-model speed.
Shares the Harmony format with GPT OSS 120B — route hard requests up without changing your integration.
High-volume classification and routing
Reasoning quality where rule-based systems fail, at a price point that survives millions of calls.
Structured extraction with strict schemas
Harmony-trained function calling and JSON output that parse the first time.
Latency-sensitive assistants
Top-15 output speed on Artificial Analysis with sub-second first tokens.
Escalation-tier architectures
Serve the routine calls here and escalate hard ones to GPT OSS 120B — same Harmony format, drop-in swap.
Repo-scale coding agents
Multi-file engineering is beyond a 3.6B-active model — use DeepSeek V4 Flash or Qwen 3.6 27B for issue-to-patch work.
Knowledge-heavy answers without retrieval
Compact models trail on knowledge-intensive tasks — pair it with RAG for factual workloads.
Image, audio, or video inputs
GPT OSS 20B is text-only. Route visual inputs to a multimodal model such as Gemma 4 26B A4B.
avg. 314tokens/s
avg. 350ms
98.7%AIME 2025 (with tools)
Competition mathematics measuring multi-step reasoning.
71.5%GPQA
Graduate-level scientific reasoning and analysis.
165tokens/s (Artificial Analysis)
Top-15 fastest of the 130 open models measured.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "openai/gpt-oss-20b", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
o4-mini-class reasoning for tool-driven coding
Tool-heavy coding at 3B-active cost
Gemma 4 quality at 4B-active speed