SourcesOpenAI
publisher · 4 sources
OpenAI
4 sources read from this publisher, newest first, with the methods each one describes.
gpt-oss-120b & gpt-oss-20b Model Card
first party release · technical report · 2025-08-08 · 41 methods
Mixture of expertsThinking modeHybrid Sliding-Window and Global AttentionGrouped-query attentionYaRNRotary Position EmbeddingBest-of-N selectionBrowser toolMXFP4 QuantizationAgentic tool-use post-trainingCBRN safety filtering of pre-training dataChain-of-thought reinforcement learningRMSNormRole-based instruction hierarchySwiGLUAdversarial fine-tuningCapture-the-Flag (CTF) evaluationCombined performance with refusal substitutionCyber range exercisesDeliberative alignmentExternal expert review of safety evaluation methodologyFine-tuning hyperparameter verificationharmony chat formatHelpful-only adversarial reinforcement learningIncremental reinforcement learningInference-time scaling plots for evaluation reportingInstruction hierarchy trainingInterleaved tool calls in chain-of-thoughtLLM autograding with expert validationno direct optimization pressure on chain-of-thoughto200k_harmony tokenizerOff-by-one attentionPaperBench hierarchical rubric evaluationPre-LNProtocolQA robustness validationProtocolQA-aligned RL training datasetsRefusal behavior quantificationSWE-bench Verified evaluationText-only multimodal benchmark alignment for comparabilityThree-configuration cyber range testing
openai/gpt-oss-120b · Hugging Face
first party release · model card · 2025-08-07 · 6 methods
OpenAI Harmony Response Format
software documentation · vendor docs · 2024-06-01 · 21 methods
Thinking modeExcluding prior thinking from multi-turn historyHarmony formatAssistant output channelsRole-based instruction hierarchyBuilt-in tool declaration in the system messageChain-of-thought in the analysis channelCommentary-channel preamblesDeveloper message formatGrammar-constrained decodingHarmony function-calling formatHarmony history end-token normalizationHarmony system message formatHarmony tool-call message formatJSON Schema response formatsOpenAI Harmony renderer (openai_harmony) that renders messages into tokensPython execution tool in chain-of-thoughtRetaining chain-of-thought across tool callsStreamableParserTool message formatTypeScript-like function schema syntax
openai/gpt-oss
first party release · code repo · 18 methods
Mixture of expertsThinking modeCUDA Graph capture and replayBrowser toolHarmony formatMXFP4 QuantizationBF16 inferenceFull chain-of-thought outputAttention memory-cost optimization in the Triton implementationBrowser-tool request cachingMXFP4 block-scaled tensor packingOptimized Triton MoE kernel with MXFP4 supportPython-tool use during chain-of-thoughtPyTorch expandable segments allocator (PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True) for loading checkpointsSampling with temperature=1.0 and top_p=1.0Stateful Python toolStateless Python tool reference implementationTensor parallelism for MoE layers