GPT OSS 20B
OpenAI's open-source 20B model with hybrid reasoning, extended thinking, efficient code generation, agentic search, and tool use. Cost-effective alternative to the 120B variant.
GPT OSS 20B is the smaller sibling of OpenAI's first open-weight language model family since GPT-2, released in August 2025 under the Apache 2.0 license. It is a 21B-parameter Mixture-of-Experts (MoE) transformer that activates only 3.6B parameters per forward pass, making it deployable on devices with as little as 16GB of memory — including edge hardware and consumer laptops. Despite its compact size, it was post-trained with the same chain-of-thought reinforcement learning pipeline as the 120B variant, delivering configurable reasoning effort (low / medium / high), full CoT visibility, tool use, and Structured Outputs.
In production, GPT OSS 20B is ideal for edge scenarios, on-device AI, and rapid iteration workflows where GPU budget is limited or latency must be minimized. It matches or exceeds OpenAI o3-mini on most benchmarks, often outperforming it on competition mathematics and health-related reasoning — remarkable for a 21B total / 3.6B active parameter model. Through TheRouter, you can access it via a managed OpenAI-compatible API without self-hosting; alternatively, the Apache 2.0 license lets you download, fine-tune, and deploy it entirely on your own infrastructure.
- • Edge and local inference — runs on 16GB devices including laptops, edge servers, and single-board computers.
- • Cost-sensitive agentic workflows — tool use, function calling, and agent loops at a fraction of the cost of larger frontier models.
- • Rapid prototyping and iteration — fast CoT with low reasoning effort, ideal for debugging prompts and agent orchestration before scaling up.
- • On-device customization — Apache 2.0 license permits unfettered fine-tuning and commercial deployment.
- • High-complexity reasoning under one shot — for tasks requiring sustained deep reasoning (e.g., hard HLE problems), prefer gpt-oss-120b in high reasoning mode.
- • Multimodal tasks — this is a text-only model. For vision or audio, use GPT-4o or GPT-4o-mini via TheRouter.
- • Latency-critical high-throughput APIs — the medium/high reasoning modes still produce CoT traces; for bare-minimum latency, pair low reasoning effort with Mistral Small 4 or GPT-4.1 Nano via TheRouter.
Modalities
Pricing Breakdown
| Type | Rate |
|---|---|
| Input | $0.0756 / 1M tokens |
| Output | $0.324 / 1M tokens |
Supported Parameters
Specifications
| Release date | 2025-08-05openai.com ↗ | verified |
| Architecture | MoE Transformer — 24 layers, 32 experts, top-4 active, 21B total / 3.6B active paramsopenai.com ↗ | verified |
| License | Apache 2.0openai.com ↗ | verified |
| Knowledge cutoff | June 2024arxiv.org ↗ | verified |
| Tokenizer | o200k_harmony (201,088 tokens)arxiv.org ↗ | verified |
| Minimum memory | 16 GB (consumer edge device deployment)openai.com ↗ | verified |
| Quantization | MXFP4 (4.25 bits/param on MoE weights)arxiv.org ↗ | verified |
| Checkpoint size | 12.8 GiBhuggingface.co ↗ | verified |
| Reasoning modes | low / medium / high (configurable in system prompt)openai.com ↗ | verified |
Benchmarks
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU High reasoning mode | 85.3%% | arxiv.org ↗ | |
AIME 2024 With tools, high reasoning mode | 96.0%% | arxiv.org ↗ | |
AIME 2025 With tools, high reasoning mode | 98.7%% | arxiv.org ↗ | |
Codeforces Elo Estimated Elo rating; high reasoning mode with terminal tool | ~2516 | openai.com ↗ | |
SWE-Bench Verified High reasoning mode | 60.7%% | arxiv.org ↗ | |
GPQA Diamond With tools, high reasoning mode | 71.5%% | arxiv.org ↗ | |
HLE With tools, high reasoning mode. HLE is an extremely difficult benchmark. | 17.3%% | arxiv.org ↗ | |
HealthBench High reasoning mode. Outperforms o3-mini on health-related reasoning despite smaller size. | 42.5%% | arxiv.org ↗ | |
Tau-Bench Retail High reasoning mode — agentic function calling evaluation | 54.8%% | arxiv.org ↗ |
API Usage Examples
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-oss-20b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'API guide
Full API reference →Chat completion
Use standard OpenAI-compatible chat completion via TheRouter. The model uses the harmony chat format; TheRouter handles the conversion automatically.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-oss-20b",
"messages": [
{"role": "user", "content": "What makes Mixture-of-Experts models efficient for edge deployment?"}
]
}'More from openai
Similar models
Cross-provider sibling modelsFrequently asked
Can I use GPT OSS 20B through TheRouter's API without self-hosting?
Yes. TheRouter routes requests to gpt-oss-20b through its managed infrastructure. You access the open-weight model via a standard OpenAI-compatible API at https://api.therouter.ai/v1, without managing your own GPU hardware. The 20B variant is particularly well-suited for rapid iteration and edge-scale workloads.
What license does GPT OSS 20B use?
Apache 2.0 — a permissive license that allows commercial use, modification, distribution, and sublicensing without copyleft restrictions or patent risk.
How does GPT OSS 20B compare to GPT OSS 120B?
GPT OSS 20B is the smaller (21B total / 3.6B active) sibling of the 120B (117B total / 5.1B active). It runs on as little as 16GB memory instead of 80GB, while the 120B achieves higher scores on hard reasoning benchmarks (HLE, HealthBench) and agentic evaluations (Tau-Bench). The 20B still matches or exceeds o3-mini on most benchmarks — a strong tradeoff for edge deployment and cost-sensitive workloads.
What hardware do I need to self-host GPT OSS 20B?
A device with 16GB of memory — such as an NVIDIA RTX 4060/4070, an Apple Mac with M-series chip and 16GB unified memory, or an edge server GPU. The quantized checkpoint is 12.8 GiB. With MXFP4 quantization and MoE architecture, only 3.6B parameters are active per forward pass, enabling efficient local inference on consumer hardware.
Fact ledger — every claim on this page traces here
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-29 | verified |
| Architecture | openai.com ↗ | 2026-05-29 | verified |
| License | openai.com ↗ | 2026-05-29 | verified |
| Knowledge cutoff | arxiv.org ↗ | 2026-05-29 | verified |
| Tokenizer | arxiv.org ↗ | 2026-05-29 | verified |
| Minimum memory | openai.com ↗ | 2026-05-29 | verified |
| Quantization | arxiv.org ↗ | 2026-05-29 | verified |
| Checkpoint size | huggingface.co ↗ | 2026-05-29 | verified |
| Reasoning modes | openai.com ↗ | 2026-05-29 | verified |
| MMLU | arxiv.org ↗ | 2026-05-29 | verified |
| AIME 2024 | arxiv.org ↗ | 2026-05-29 | verified |
| AIME 2025 | arxiv.org ↗ | 2026-05-29 | verified |
| Codeforces Elo | openai.com ↗ | 2026-05-29 | to verify |
| SWE-Bench Verified | arxiv.org ↗ | 2026-05-29 | verified |
| GPQA Diamond | arxiv.org ↗ | 2026-05-29 | verified |
| HLE | arxiv.org ↗ | 2026-05-29 | verified |
| HealthBench | arxiv.org ↗ | 2026-05-29 | verified |
| Tau-Bench Retail | arxiv.org ↗ | 2026-05-29 | verified |
| Can I use GPT OSS 20B through TheRouter's API without self-hosting? | openai.com ↗ | 2026-05-29 | to verify |
| What license does GPT OSS 20B use? | openai.com ↗ | 2026-05-29 | to verify |
| How does GPT OSS 20B compare to GPT OSS 120B? | openai.com ↗ | 2026-05-29 | to verify |
| What hardware do I need to self-host GPT OSS 20B? | github.com ↗ | 2026-05-29 | to verify |