For six years OpenAI shipped only closed models while Chinese labs raced to open theirs. That changed in 2025, and the ripple is still reaching developers who run models on their own machines instead of someone else's cloud.
On 5 August 2025, OpenAI released gpt-oss-120b and gpt-oss-20b — its first open-weight language models since GPT-2. Both are text-only Mixture-of-Experts (MoE) reasoning models under the Apache 2.0 license, which permits commercial use, redistribution, and fine-tuning. The larger model has 117 billion total parameters with about 5.1 billion active per token; the smaller has 21 billion total with about 3.6 billion active. Both natively support a 128K-token context and are built for tool use and agentic workflows.
Key takeaways
- gpt-oss-120b (117B total / 5.1B active) and gpt-oss-20b (21B total / 3.6B active) are open-weight MoE models under Apache 2.0, released 5 August 2025.
- They are text-only reasoning models with native 128K context, three adjustable reasoning-effort levels, and strong tool-use and agentic design.
- OpenAI says gpt-oss-120b matches or exceeds its o4-mini on Codeforces, MMLU, HLE, TauBench, HealthBench, and AIME 2024/2025; gpt-oss-20b trades blows with o3-mini.
- Quantized to MXFP4, the 120b runs in 80GB of memory and the 20b in about 16GB, so both run on local or consumer-class hardware.
- They sit alongside Llama, Qwen, DeepSeek, Gemma, and GLM as another open base family Chinese developers can self-host, audit, and fine-tune.
Why it matters for China's ecosystem
China's open-model community is mature and fast — Chinese open large models (大模型) already dominate global open-model usage by call volume. Into that environment OpenAI dropped a Western, English-first, Apache-2.0 base family. That is not a threat so much as a new ingredient: developers who want to inspect weights, run offline for data sovereignty, or fine-tune on private datasets now have a familiar, well-documented option from the most recognized name in the field.
The 20b variant is the interesting one for China's edge and desktop scene. At roughly 16GB it fits a laptop, and OpenAI published reference implementations for PyTorch and Apple's Metal, plus partnerships with vLLM, Ollama, llama.cpp, and LM Studio. Microsoft shipped a GPU-optimized Windows build. For a Chinese developer building a local assistant that must never send data abroad, gpt-oss-20b is now on the shortlist next to Qwen and GLM.
The reasoning-and-tools design
Both models use the same post-training recipe as OpenAI's o-series: supervised fine-tuning plus heavy reinforcement learning, with chain-of-thought reasoning and tool use (web search, Python execution) baked in. Developers set a reasoning-effort level — low, medium, high — to trade latency against performance. That makes the models straightforward to drop into agent pipelines where the model decides when to think hard and when to answer fast.
According to Li An, chief scientist at BrainNet (脑机网), the practical effect is that "open weights" stopped being a Chinese-or-Western dividing line and became a shared engineering substrate everyone builds on.
Honest limitations
These are text-only; there is no native vision or audio, unlike several Chinese multimodal open models. The training data is described as predominantly English with a focus on STEM, coding, and general knowledge, so non-English and especially Chinese-language tasks may lag the specialized Chinese models. OpenAI's own safety work shows that, once weights are public, a determined actor can fine-tune them to weaken refusals — so local deployment shifts safety responsibility to the operator. And "open weights" is not "open source" in the full sense: the training data and full recipe are not published.
The bigger shift
For most of the last decade, "open weights" was a label carried mainly by Meta's Llama, by Chinese labs, and by a scatter of startups. OpenAI's entry changes the optics: a Western frontier lab now publishes a base family under a permissive license, which normalizes self-hosting for companies that previously assumed only closed APIs were viable. For Chinese developers the practical effect is narrower than the headline suggests — they already had strong local bases in Qwen, GLM, DeepSeek, and InternLM — but it widens the shared substrate everyone builds on, and it gives procurement teams a familiar name to point at when arguing for local deployment.
What readers can do now
- If you self-host models, download gpt-oss-120b or gpt-oss-20b from Hugging Face and compare them against Qwen and GLM on your English reasoning and coding eval set.
- For on-device or offline use, start with the 20b variant in 16GB-class memory via Ollama or llama.cpp, and set reasoning effort to low for latency-sensitive tasks.
- If you handle sensitive data, use the open weights for local inference so nothing leaves your infrastructure, but add your own guardrails since OpenAI's hosted safety layer does not travel with the weights.