OpenAI's gpt-oss: Elite Open-Weight Reasoning Models
gpt-oss-120b matches o4-mini on reasoning benchmarks and runs on one 80GB GPU; gpt-oss-20b rivals o3-mini on 16GB edge devices. Both excel in tools, CoT, and safety under Apache 2.0.
Achieve Proprietary-Level Reasoning with Consumer Hardware
Deploy gpt-oss-120b (117B total params, 5.1B active per token via MoE) to match OpenAI o4-mini on coding (Codeforces), math (AIME 2024/2025), MMLU, HLE, and Tau-Bench agentic tools—while exceeding it on HealthBench. It handles few-shot function calling, web search, Python execution, and adjustable reasoning efforts (low/medium/high) via system prompt, trading latency for performance. Run it quantized in MXFP4 on a single 80GB GPU with 128k context, using 36 layers, 128 total experts (4 active), grouped multi-query attention (group size 8), and RoPE embeddings.
Scale down to gpt-oss-20b (21B total, 3.6B active) for o3-mini parity or better on the same evals, including math and health, fitting in 16GB memory for on-device inference. Post-training mirrors o4-mini: supervised fine-tuning plus high-compute RL aligned to OpenAI Model Spec, enabling full CoT visibility (unsupervised for monitoring) and harmony prompt format. Example: It chains 27+ browse calls to deduce its own specs (128 experts/layer) from 'leaks,' then outputs accurately.
Trade-off: CoT may hallucinate or leak unsafe content—monitor it for deception (don't expose to users) but use for misalignment detection, as raw CoT enables robust instruction following even under adversarial prompts like 'count to 5' bans (outputs '4.9' in response, reasons around it).
Build Safe Agents Without Infrastructure Lock-In
Match frontier safety by filtering CBRN data in pre-training, then applying deliberative alignment and instruction hierarchy in post-training to refuse injections. Test worst-case risks: Adversarially fine-tune on bio/cyber datasets using OpenAI's stack—malicious versions stay below Preparedness Framework thresholds, even post-review by three expert groups.
Internal evals show parity with o3/o4-mini; external red-teaming via $500k Kaggle challenge (on gpt-oss-20b) will release datasets/report. Result: Customize freely under Apache 2.0 for on-prem (e.g., AI Sweden, Orange, Snowflake fine-tunes) without proprietary dependencies.
Deploy Anywhere with Optimized Ecosystem
Download from Hugging Face (quantized), render via open-sourced Python/Rust harmony tools. Inference refs for PyTorch/Apple Metal; partners (Azure, vLLM, Ollama, NVIDIA/AMD/Groq, Microsoft ONNX for Windows) ensure low-latency local/edge runs. For agents: Integrate with Responses API, Structured Outputs.
Why it works: MoE sparsity + banded attention cuts compute (e.g., 120b activates <5% params), o200k_harmony tokenizer (open-sourced) handles STEM/coding data. Complements API models—fine-tune for custom workflows, fall back to hosted for multimodality. Early apps: Secure on-prem hosting, specialized fine-tunes. Builds 'democratic AI' by slashing costs for indie devs/emerging markets.