{ "schema_version": 1, "asset_policy": "image_preinstalled", "target": "spacemit-k3-32g-minicpm-chat-qwen-summary", "model_set_version": "2026.08.01-minicpm-chat-qwen-summary", "description": "Mixed K3 model combination: MiniCPM5-1B Q4_K_M is the hot llm-chat model and Qwen3-30B-A3B Q4_0 is the deferred llm-summary model; retrieval, OCR, ASR, TTS, and compact VLM match the MiniCPM profile.", "runtime_memory_policy": { "default_hot_models": [ "embedding-int8", "llm-chat" ], "feature_cold_models": [ "reranker-int8", "llm-summary", "ocr-det", "ocr-rec", "vad-x100", "asr-sherpa-x100", "asr-whisper", "tts-default", "vlm" ], "stress_only_models": [], "default_cold_models": [], "soft_min_available_gb": 6.0, "hard_min_available_gb": 2.0, "trim_after_feature_task": true }, "packages": [ { "package": "k3-scheduler-model-ocr-ppocrv5-mobile", "title": "PP-OCRv5 mobile OCR assets", "role": "document_ocr", "runtime_tier": "hot_path", "default_install": true, "depends": [ "spacemit-onnxruntime" ], "scheduler_models": [ "ocr-det", "ocr-rec" ], "description": "PPOCR detection and recognition ONNX files used by document upload and OCR stress flows. K3 defaults run the scheduler OCR workers through the SpacemiT ONNX Runtime EP.", "assets": [ { "type": "file", "source_candidates": [ "vision/ppocr/PP-OCRv5_mobile_det.onnx", "spacemit-ai/vision/ppocr/PP-OCRv5_mobile_det.onnx" ], "target": "/root/models/spacemit-ai/vision/ppocr/PP-OCRv5_mobile_det.onnx" }, { "type": "file", "source_candidates": [ "vision/ppocr/PP-OCRv5_mobile_rec.onnx", "spacemit-ai/vision/ppocr/PP-OCRv5_mobile_rec.onnx" ], "target": "/root/models/spacemit-ai/vision/ppocr/PP-OCRv5_mobile_rec.onnx" }, { "type": "file", "optional": false, "source_candidates": [ "vision/ppocr/ppocrv5_dict.txt", "spacemit-ai/vision/ppocr/ppocrv5_dict.txt" ], "target": "/root/models/spacemit-ai/vision/ppocr/ppocrv5_dict.txt", "validation": { "kind": "ppocrv5_dictionary", "encoding": "utf-8", "line_count": 18383, "recognition_classes": 18385 } } ], "expected_paths": [ "/root/models/spacemit-ai/vision/ppocr/PP-OCRv5_mobile_det.onnx", "/root/models/spacemit-ai/vision/ppocr/PP-OCRv5_mobile_rec.onnx", "/root/models/spacemit-ai/vision/ppocr/ppocrv5_dict.txt" ] }, { "package": "k3-scheduler-model-sherpa-x100-audio", "title": "rv-sherpa-onnx X100 VAD and streaming ASR assets", "role": "realtime_audio_frontend", "runtime_tier": "feature_cold", "default_install": true, "depends": [ "rv-sherpa-onnx-runtime" ], "scheduler_models": [ "vad-x100", "asr-sherpa-x100" ], "composite_models": [ { "name": "asr-diarization", "components": [ "vad-x100", "asr-sherpa-x100" ], "strategy": "vad_segmented_streaming_asr", "output_schema": "meeting_segments.v1", "quality_scope": "VAD-driven speech segmentation plus streaming ASR transcript segments. Speaker identity clustering requires a future speaker embedding package." } ], "description": "K3 sherpa audio frontend profile validated on 2026-07-18. X100 VAD and streaming ASR are the production-relevant low-latency audio path. The asr-diarization capability is delivered as a VAD+ASR composite, not as a standalone segmentation.onnx artifact.", "assets": [ { "type": "file", "source_candidates": [ "silero-vad/onnx/model.onnx" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/vad-silero/model.onnx" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/tokens.txt" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/tokens.txt" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/bpe.model" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/bpe.model" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/bpe.vocab" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/bpe.vocab" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/encoder-epoch-99-avg-1.int8.onnx" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/encoder-epoch-99-avg-1.int8.onnx" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/decoder-epoch-99-avg-1.int8.onnx" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/decoder-epoch-99-avg-1.int8.onnx" }, { "type": "file", "source_candidates": [ "sherpa-onnx/sherpa-onnx-streaming-zipformer-bilingual-zh-en-2023-02-20/joiner-epoch-99-avg-1.int8.onnx" ], "target": "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/joiner-epoch-99-avg-1.int8.onnx" } ], "expected_paths": [ "/opt/rv-sherpa-onnx/models/candidates/current/vad-silero/model.onnx", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/tokens.txt", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/bpe.model", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/bpe.vocab", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/encoder-epoch-99-avg-1.int8.onnx", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/decoder-epoch-99-avg-1.int8.onnx", "/opt/rv-sherpa-onnx/models/candidates/current/asr-streaming-zipformer/joiner-epoch-99-avg-1.int8.onnx" ], "bench_evidence": { "source": "k3-rv-sherpa-audio-e2e-20260718", "host": "192.168.100.140", "provider": "x100", "vad_p95_decision_ms": 0.718378, "asr_p95_first_stable_ms": 382.791704, "asr_streaming_rtf": 0.306315, "quality_scope": "functional_and_latency; authoritative ASR transcript corpus still required" } }, { "package": "k3-scheduler-model-asr-qwen3-0p6b", "title": "Qwen3-ASR 0.6B frontend assets", "role": "meeting_asr", "runtime_tier": "feature_cold", "default_install": true, "depends": [ "spacemit-onnxruntime" ], "scheduler_models": [ "asr-whisper" ], "description": "Qwen3-ASR 0.6B package used for meeting-audio frontend pressure. The current scheduler worker uses the encoder frontend ONNX stage.", "assets": [ { "type": "tar_gz", "source_candidates": [ "vlm/qwen3-asr-0.6B.tar.gz", "asr/qwen3-asr-0.6B-dynq-q40.tar.gz" ], "target_dir": "/root/models/spacemit-ai/vlm" } ], "expected_paths": [ "/root/models/spacemit-ai/vlm/qwen3-asr-0.6B/Qwen3-ASR-0.6B-encoder-frontend.dynq.onnx" ] }, { "package": "k3-scheduler-model-tts-matcha-zh-baker", "title": "Sherpa Matcha Chinese product TTS assets", "role": "speech_synthesis", "runtime_tier": "feature_cold", "default_install": true, "depends": [ "rv-sherpa-onnx-runtime" ], "scheduler_models": [ "tts-default" ], "description": "Official sherpa-onnx Matcha Chinese acoustic model, Vocos 22 kHz vocoder, lexicon, dictionary, tokens, and normalization FSTs used by the K3 C++ audio proxy.", "assets": [ { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/model-steps-3.onnx"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/acoustic_model.onnx" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/vocos-22khz-univ.onnx"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/vocoder.onnx" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/lexicon.txt"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/lexicon.txt" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/tokens.txt"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/tokens.txt" }, { "type": "directory", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/dict"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/dict" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/phone.fst"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/phone.fst" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/date.fst"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/date.fst" }, { "type": "file", "source_candidates": ["sherpa-onnx/tts/matcha-icefall-zh-baker/number.fst"], "target": "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/number.fst" } ], "expected_paths": [ "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/acoustic_model.onnx", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/vocoder.onnx", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/lexicon.txt", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/tokens.txt", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/dict", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/phone.fst", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/date.fst", "/opt/rv-sherpa-onnx/models/candidates/current/tts-matcha/number.fst" ], "source_evidence": { "acoustic_archive_url": "https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/matcha-icefall-zh-baker.tar.bz2", "acoustic_archive_sha256": "20de2ec034b55562609d6362771c934905dfe11d0f41ec103d593427ad9a7efb", "vocoder_url": "https://github.com/k2-fsa/sherpa-onnx/releases/download/vocoder-models/vocos-22khz-univ.onnx", "vocoder_sha256": "0574a135aa1db2de6e181050db2ec528496cacd4a4701fc5d7faf9f9804c0081", "validation_scope": "source integrity and host-side asset contract; K3 product audio E2E remains pending while the device is offline" } }, { "package": "k3-scheduler-model-tts-ort-contract", "title": "ORT token_ids_v1 TTS generation contract", "role": "speech_synthesis_ort_contract", "runtime_tier": "feature_cold", "default_install": false, "depends": [ "spacemit-onnxruntime" ], "scheduler_models": [], "description": "Synthetic token_ids_v1 ONNX fixture retained for host contract tests. It is not a product TTS model and is excluded from the default image profile.", "assets": [ { "type": "base64_file", "source_candidates": [ "tests/fixtures/tts/synthetic-token-ids-v1.onnx.b64" ], "target": "/root/models/spacemit-ai/tts/synthetic-token-ids-v1.onnx", "sha256": "dc22dcc1ee22a8a203fcd27f701f6ff7e985fa3352c465d204ca324ad84f24f6" }, { "type": "file", "source_candidates": [ "tests/fixtures/tts/tokens.json" ], "target": "/root/models/spacemit-ai/tts/tokens.json" } ], "expected_paths": [ "/root/models/spacemit-ai/tts/synthetic-token-ids-v1.onnx", "/root/models/spacemit-ai/tts/tokens.json" ], "bench_evidence": { "source": "tests/tts_ort_worker_fixture.py", "provider": "ort_cpu", "schema": "speech_audio.v1", "generation_contract": true, "sample_rate_hz": 8000, "duration_ms": 100, "quality_scope": "scheduler/ORT TTS contract only; final voice model selection remains bench-owned" } }, { "package": "k3-scheduler-model-kb-bge-small-zh-rerank", "title": "BGE Chinese embedding and reranker assets", "role": "knowledge_retrieval", "runtime_tier": "hot_path", "default_install": true, "depends": [ "llama.cpp-tools-spacemit" ], "scheduler_models": [ "embedding-int8", "reranker-int8" ], "description": "BGE Chinese embedding and BGE reranker GGUF files served through llama.cpp-tools-spacemit.", "assets": [ { "type": "file", "source_candidates": [ "embed/Bge-Small-Zh-V1.5-Q4_K_M.gguf", "spacemit-ai/embed/Bge-Small-Zh-V1.5-Q4_K_M.gguf" ], "target": "/root/models/spacemit-ai/embed/Bge-Small-Zh-V1.5-Q4_K_M.gguf" }, { "type": "file", "source_candidates": [ "rerank/Bge-Reranker-V2-M3-Q4_0.gguf", "spacemit-ai/rerank/Bge-Reranker-V2-M3-Q4_0.gguf" ], "target": "/root/models/spacemit-ai/rerank/Bge-Reranker-V2-M3-Q4_0.gguf" } ], "expected_paths": [ "/root/models/spacemit-ai/embed/Bge-Small-Zh-V1.5-Q4_K_M.gguf", "/root/models/spacemit-ai/rerank/Bge-Reranker-V2-M3-Q4_0.gguf" ] }, { "package": "k3-scheduler-model-vlm-qwen35-2b", "title": "Qwen3.5-2B compact VLM assets", "role": "compact_vlm", "runtime_tier": "hot_asset_with_feature_cold_worker", "default_install": true, "depends": [ "llama.cpp-tools-spacemit" ], "scheduler_models": [ "vlm" ], "description": "Qwen3.5-2B SMT VLM package kept in the MiniCPM K3 combo for document visual extraction; MiniCPM replaces only text LLM roles in this profile.", "assets": [ { "type": "tar_gz", "source_candidates": [ "vlm/Qwen3.5-2B.tar.gz" ], "target_dir": "/root/models/spacemit-ai/vlm" } ], "expected_paths": [ "/root/models/spacemit-ai/vlm/Qwen3.5-2B/qwen3_5_2b-text-q41.gguf" ] }, { "package": "k3-scheduler-model-llm-minicpm-chat", "title": "MiniCPM5-1B K3 chat LLM asset", "role": "minicpm_chat_llm", "runtime_tier": "hot_path_chat", "default_install": true, "depends": [ "llama.cpp-tools-spacemit" ], "scheduler_models": [ "llm-chat" ], "description": "MiniCPM5-1B Q4_K_M handles hot RAG answer/chat in the mixed K3 profile. Summary is served by the separate Qwen3-30B-A3B package.", "assets": [ { "type": "file", "source_candidates": [ "/root/local-ai-bench/minicpm-probe/models/llm/MiniCPM5-1B/MiniCPM5-1B-Q4_K_M.gguf", "llm/MiniCPM5-1B/MiniCPM5-1B-Q4_K_M.gguf" ], "target": "/root/models/spacemit-ai/llm/MiniCPM5-1B/MiniCPM5-1B-Q4_K_M.gguf" } ], "expected_paths": [ "/root/models/spacemit-ai/llm/MiniCPM5-1B/MiniCPM5-1B-Q4_K_M.gguf" ], "bench_evidence": { "source": "vlm-llm-benchmark/docs/minicpm-k3-rag-replacement.zh.md", "model_root": "/root/local-ai-bench/minicpm-probe/models/llm", "install_root": "/root/models/spacemit-ai/llm", "host_verified": "192.168.100.233", "MiniCPM5-1B-Q4_K_M": { "role": "rag_answer_chat", "prompt_eval_tok_s": 330.645, "generation_tok_s": 35.682 }, "retrieval_policy": "keep_bge_embedding_and_reranker_until_minicpm_retrieval_artifacts_are_converted" } }, { "package": "k3-scheduler-model-llm-qwen3-30b-a3b", "title": "Qwen3-30B-A3B heavy LLM asset", "role": "heavy_summary_llm", "runtime_tier": "feature_cold_summary", "default_install": true, "depends": [ "llama.cpp-tools-spacemit" ], "split_deb": true, "split_chunk_bytes": 6000000000, "scheduler_models": [ "llm-summary" ], "description": "Qwen3-30B-A3B Q4_0 asset used as llm-summary in the mixed K3 profile while MiniCPM5-1B remains the hot chat/RAG answer model.", "assets": [ { "type": "file", "source_candidates": [ "llm/Qwen3-30B-A3B-Q4_0.gguf", "spacemit-ai/llm/Qwen3-30B-A3B-Q4_0.gguf" ], "target": "/root/models/spacemit-ai/llm/Qwen3-30B-A3B-Q4_0.gguf" } ], "expected_paths": [ "/root/models/spacemit-ai/llm/Qwen3-30B-A3B-Q4_0.gguf" ] } ], "meta_packages": [ { "package": "k3-scheduler-models-kb-default", "title": "Default K3 MiniCPM chat plus Qwen summary model set", "include": [ "k3-scheduler-model-ocr-ppocrv5-mobile", "k3-scheduler-model-sherpa-x100-audio", "k3-scheduler-model-asr-qwen3-0p6b", "k3-scheduler-model-tts-matcha-zh-baker", "k3-scheduler-model-kb-bge-small-zh-rerank", "k3-scheduler-model-vlm-qwen35-2b", "k3-scheduler-model-llm-minicpm-chat", "k3-scheduler-model-llm-qwen3-30b-a3b" ], "description": "Installs OCR, ASR frontend, TTS generation contract, BGE retrieval, compact VLM, MiniCPM chat, and Qwen3-30B summary assets for the mixed K3 combo." }, { "package": "k3-scheduler-models-full-stress", "title": "Full K3 mixed MiniCPM/Qwen scheduler stress model set", "include": [ "k3-scheduler-model-ocr-ppocrv5-mobile", "k3-scheduler-model-sherpa-x100-audio", "k3-scheduler-model-asr-qwen3-0p6b", "k3-scheduler-model-tts-matcha-zh-baker", "k3-scheduler-model-tts-ort-contract", "k3-scheduler-model-kb-bge-small-zh-rerank", "k3-scheduler-model-vlm-qwen35-2b", "k3-scheduler-model-llm-minicpm-chat", "k3-scheduler-model-llm-qwen3-30b-a3b" ], "description": "Installs the mixed knowledge-base model assets plus optional sherpa X100 audio frontend." } ] }