{"id":"model.alibaba.qwen.qwen3-30b-a3b","name":"Qwen3-30B-A3B","family":"Qwen3","type":"frontier","modalities":["text"],"capabilities":["agent"],"openness":{"weights":"open_weights","code":"not_disclosed","data":"partially_disclosed","license":null,"deployment":"self_hostable","api":null},"keySpecs":{"contextWindowTokens":null,"extendedContextWindowTokens":null,"totalParameters":null,"activeParameters":null,"architectureType":null,"baseModel":null,"inputPriceUsdPerMtok":null,"outputPriceUsdPerMtok":null,"concurrencyLimitRequests":null},"organization":{"id":"org.alibaba-qwen","name":"阿里千问 / Qwen","websiteUrl":"https://qwen.ai/"},"brand":{"id":"brand.alibaba-qwen","name":"Qwen","websiteUrl":"https://qwen.ai"},"product":{"id":"product.alibaba-qwen.models","name":"Qwen Models","websiteUrl":"https://qwen.readthedocs.io/en/latest/"},"origin":"domestic","versions":[{"id":"version.alibaba.qwen.qwen3-30b-a3b","name":"Qwen3-30B-A3B","externalModelId":"Qwen/Qwen3-30B-A3B","kind":"date","releaseDate":"2025-04-29","datePrecision":"day","publicAvailableAt":null,"lifecycleStatus":"active","knowledgeCutoff":null,"claims":[{"id":"claim.version.alibaba.qwen.qwen3-30b-a3b.capability.agent","fieldKey":"capability_tags","value":{"value":["agent"]},"rawValue":"Qwen3 supports tool calling and agent capabilities","ingestionOrigin":"curated_official","unit":null,"condition":"Explicit public claim projection from agent; this is a documented capability tag, not a benchmark or quality conclusion.","status":"approved","confidence":"high","reviewedBy":"system:capability-projection","reviewedAt":"2026-08-05 09:26:02.481569+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.docs","title":"Qwen documentation","url":"https://qwen.readthedocs.io/en/latest/","accessedAt":"2026-09-17 15:20:45.271+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:00.621+08"},"evidence":{"id":"evidence.alibaba.qwen.docs.context","locator":"Key Concepts and Qwen3-2507 documentation","excerpt":"Qwen documents 119 languages and dialects for Qwen3. For Qwen3-2507, the maximum assistant message length is 81,920 tokens for thinking models and 16,384 tokens for instruct models; Qwen3-2507 supports 256K context and can be extended to 1M tokens."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.deployment","fieldKey":"deployment_options","value":["Transformers","vLLM","SGLang","Ollama","LM Studio","MLX-LM","llama.cpp","KTransformers"],"rawValue":"Official documentation and model card list Transformers, vLLM, SGLang and local deployment tools.","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:17:18.866955+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.docs","title":"Qwen documentation","url":"https://qwen.readthedocs.io/en/latest/","accessedAt":"2026-09-17 15:20:45.271+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:00.621+08"},"evidence":{"id":"evidence.alibaba.qwen.docs.context","locator":"Key Concepts and Qwen3-2507 documentation","excerpt":"Qwen documents 119 languages and dialects for Qwen3. For Qwen3-2507, the maximum assistant message length is 81,920 tokens for thinking models and 16,384 tokens for instruct models; Qwen3-2507 supports 256K context and can be extended to 1M tokens."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.languages","fieldKey":"language_count","value":{"unit":"languages and dialects","value":119},"rawValue":"119 languages and dialects","ingestionOrigin":"curated_official","unit":"languages","condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.models","locator":"Model tables and release metadata","excerpt":"Qwen states that Qwen3 was released on April 29, 2025, open-weights two MoE models and six dense models under Apache 2.0. The model tables list layers, attention heads, experts, and 128K context for the release family."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.post-training-methods","fieldKey":"post_training_methods","value":{"value":"Long CoT cold start; reasoning-based RL; thinking-mode fusion; general RL."},"rawValue":"Four-stage post-training pipeline","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.training","locator":"Pre-training and Post-training sections","excerpt":"Qwen states that Qwen3 uses approximately 36 trillion pre-training tokens across 119 languages and dialects. The pre-training process has three stages; post-training has four stages: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general reinforcement learning."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.pretraining-tokens","fieldKey":"pretraining_tokens","value":{"unit":"T","value":36},"rawValue":"Approximately 36 trillion pre-training tokens","ingestionOrigin":"curated_official","unit":"T","condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.training","locator":"Pre-training and Post-training sections","excerpt":"Qwen states that Qwen3 uses approximately 36 trillion pre-training tokens across 119 languages and dialects. The pre-training process has three stages; post-training has four stages: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general reinforcement learning."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.release-date","fieldKey":"release_date","value":{"value":"2025-04-29","precision":"day"},"rawValue":"April 29, 2025","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.models","locator":"Model tables and release metadata","excerpt":"Qwen states that Qwen3 was released on April 29, 2025, open-weights two MoE models and six dense models under Apache 2.0. The model tables list layers, attention heads, experts, and 128K context for the release family."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.thinking","fieldKey":"thinking_modes","value":{"value":"hybrid: thinking and non-thinking"},"rawValue":"Seamless switching between thinking and non-thinking modes","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.models","locator":"Model tables and release metadata","excerpt":"Qwen states that Qwen3 was released on April 29, 2025, open-weights two MoE models and six dense models under Apache 2.0. The model tables list layers, attention heads, experts, and 128K context for the release family."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.tool-calls","fieldKey":"tool_calls","value":{"value":true},"rawValue":"Qwen3 supports tool calling and agent capabilities","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.docs","title":"Qwen documentation","url":"https://qwen.readthedocs.io/en/latest/","accessedAt":"2026-09-17 15:20:45.271+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:00.621+08"},"evidence":{"id":"evidence.alibaba.qwen.docs.context","locator":"Key Concepts and Qwen3-2507 documentation","excerpt":"Qwen documents 119 languages and dialects for Qwen3. For Qwen3-2507, the maximum assistant message length is 81,920 tokens for thinking models and 16,384 tokens for instruct models; Qwen3-2507 supports 256K context and can be extended to 1M tokens."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.training-data-transparency","fieldKey":"training_data_transparency","value":{"value":"partially_disclosed"},"rawValue":"Qwen discloses aggregate token count, language coverage and data construction methods, not the full dataset.","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:17:18.866955+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.training","locator":"Pre-training and Post-training sections","excerpt":"Qwen states that Qwen3 uses approximately 36 trillion pre-training tokens across 119 languages and dialects. The pre-training process has three stages; post-training has four stages: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general reinforcement learning."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.training-methods","fieldKey":"training_methods","value":{"value":"Three-stage pretraining: S1 over 30T tokens at 4K context; S2 adds 5T knowledge-intensive data; S3 uses high-quality long-context data to extend context to 32K."},"rawValue":"Three-stage pretraining","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.training","locator":"Pre-training and Post-training sections","excerpt":"Qwen states that Qwen3 uses approximately 36 trillion pre-training tokens across 119 languages and dialects. The pre-training process has three stages; post-training has four stages: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general reinforcement learning."}},{"id":"claim.version.alibaba.qwen.qwen3-30b.weights","fieldKey":"weights_availability","value":{"value":"open_weights"},"rawValue":"Qwen describes this model as open-weighted","ingestionOrigin":"curated_official","unit":null,"condition":null,"status":"approved","confidence":"high","reviewedBy":"system:curated_official","reviewedAt":"2026-08-03 22:13:04.249287+08","effectiveFrom":null,"effectiveTo":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/","accessedAt":"2026-09-17 15:20:54.384+08","robotsStatus":"allowed","collectionEligibility":"eligible","collectionCheckedAt":"2026-09-17 15:17:03.59+08"},"evidence":{"id":"evidence.alibaba.qwen.qwen3.blog.models","locator":"Model tables and release metadata","excerpt":"Qwen states that Qwen3 was released on April 29, 2025, open-weights two MoE models and six dense models under Apache 2.0. The model tables list layers, attention heads, experts, and 128K context for the release family."}}],"benchmarks":[]}],"dataUpdatedAt":"2026-08-05 09:26:02.481569+08","latestVersion":{"id":"version.alibaba.qwen.qwen3-30b-a3b","name":"Qwen3-30B-A3B","externalModelId":"Qwen/Qwen3-30B-A3B","status":"active"},"dataQuality":{"approved":11,"pending":0,"stale":11,"conflicted":0},"availabilityRegions":[],"availabilityChannels":[],"availabilityScopes":[],"decisionReadiness":{"overall":"incomplete","evaluatedAt":"2026-11-01T12:11:54.847Z","tracks":[{"key":"api","applicable":false,"grade":"incomplete","verifiedAt":null,"satisfiedFields":[],"missingFields":[],"missingEvidenceFields":[],"staleFields":[],"conflictedFields":[],"ineligibleSourceFields":[],"reasons":[],"fields":[]},{"key":"self_hosted","applicable":true,"grade":"incomplete","verifiedAt":null,"satisfiedFields":["release_date","capability_tags","deployment_options"],"missingFields":["lifecycle_status"],"missingEvidenceFields":[],"staleFields":["context_window_tokens","weights_availability","model_license","total_parameters"],"conflictedFields":[],"ineligibleSourceFields":[],"reasons":[{"fieldKey":"context_window_tokens","code":"stale"},{"fieldKey":"weights_availability","code":"stale"},{"fieldKey":"model_license","code":"stale"},{"fieldKey":"total_parameters","code":"stale"},{"fieldKey":"lifecycle_status","code":"missing"}],"fields":[{"fieldKey":"release_date","status":"satisfied","verifiedAt":"2026-09-17 15:20:54.384+08","expiresAt":"2026-12-16T07:20:54.384Z"},{"fieldKey":"lifecycle_status","status":"missing","verifiedAt":null,"expiresAt":null},{"fieldKey":"capability_tags","status":"satisfied","verifiedAt":"2026-09-17 15:20:45.271+08","expiresAt":"2026-12-16T07:20:45.271Z"},{"fieldKey":"deployment_options","status":"satisfied","verifiedAt":"2026-09-17 15:20:45.271+08","expiresAt":"2026-12-16T07:20:45.271Z"},{"fieldKey":"context_window_tokens","status":"stale","verifiedAt":null,"expiresAt":null},{"fieldKey":"weights_availability","status":"stale","verifiedAt":"2026-09-17 15:20:54.384+08","expiresAt":"2026-10-17T07:20:54.384Z"},{"fieldKey":"model_license","status":"stale","verifiedAt":null,"expiresAt":null},{"fieldKey":"total_parameters","status":"stale","verifiedAt":null,"expiresAt":null}]}]},"timeline":[{"id":"timeline.version.version.alibaba.qwen.qwen3-30b-a3b","kind":"version","subjectId":"version.alibaba.qwen.qwen3-30b-a3b","date":"2025-04-29","datePrecision":"day","title":"Qwen3-30B-A3B","summary":"模型版本 · active","eventType":"version","lifecycleStatus":"active","source":null},{"id":"event.alibaba.qwen.qwen3-30b.release","kind":"event","subjectId":"model.alibaba.qwen.qwen3-30b-a3b","date":"2025-04-29","datePrecision":"day","title":"Qwen3-30B-A3B 发布","summary":"Qwen 官方发布 Qwen3-30B-A3B 小型 MoE 模型。","eventType":"release","lifecycleStatus":null,"source":{"id":"source.alibaba.qwen.qwen3.blog","title":"Qwen3: Think Deeper, Act Faster","url":"https://qwenlm.github.io/blog/qwen3/"}}],"availability":[],"enterpriseCases":[],"useCases":[],"deploymentProfiles":[],"learningResources":[],"trainingProfiles":[{"id":"training-profile.llamafactory.qwen3-30b-a3b.887b850","modelId":"model.alibaba.qwen.qwen3-30b-a3b","modelName":"Qwen3-30B-A3B","versionId":"version.alibaba.qwen.qwen3-30b-a3b","versionName":"Qwen3-30B-A3B","tool":"llamafactory","toolRef":"887b850813d80a8d893ec56703d5b3a0eec46de6","supportStatus":"supported","template":"qwen3","trainingStages":["sft","dpo","kto","orpo","simpo"],"tuningMethods":["full","freeze","lora","qlora"],"modalities":["text"],"modelLocator":"YOUR_QWEN3_30B_A3B_MODEL_PATH","notes":"README 的 Qwen3 模型表与 qwen3 模板说明该系列适配；本档案不替代模型权重、许可证、显存和依赖核验。","verifiedAt":"2026-08-09 00:00:00+08","source":{"id":"source.llamafactory.readme.2026-08-09","title":"LlamaFactory README（固定 887b850 工具快照）","url":"https://github.com/hiyouga/LlamaFactory/blob/887b850813d80a8d893ec56703d5b3a0eec46de6/README_zh.md"},"evidence":{"id":"evidence.llamafactory.readme.models.2026-08-09","locator":"README：模型表、模板说明与训练方法","excerpt":"LlamaFactory README names DeepSeek 3-3.2 with template deepseek3, Llama 4 with template llama4, Qwen3 with templates qwen3/qwen3_nothink, and lists pretraining, SFT, reward-model, PPO, DPO, KTO, ORPO and SimPO training methods. It states that instruct/chat models must use their corresponding template."}}],"publication":{"approvedClaimsOnly":true,"note":"页面只展示已审核 Claim；待审核数据不会被当作公开事实。"}}