{
  "schema_version": "chinaapi.open-model-deployment.v0.2",
  "status": "published",
  "as_of": "2026-08-03",
  "scope": {
    "release_window_start": "2025-08-03",
    "release_window_end": "2026-08-03",
    "inclusion_rule": "Official weights are downloadable and at least one executable self-hosting path is documented.",
    "exclusions": [
      "API-only releases",
      "announced but unreleased weights",
      "community-only ports",
      "research checkpoints without a serving path"
    ]
  },
  "evidence_levels": {
    "A": "ChinaAPI reproduced",
    "B": "inference-framework official validated recipe",
    "C": "model-vendor documented configuration",
    "D": "third-party reproduction",
    "E": "capacity estimate only"
  },
  "scenario_evidence": {
    "vendor_positioning": "Capability or use case explicitly described by the model vendor.",
    "public_benchmark": "Capability supported by a named public benchmark or published evaluation; not reproduced by ChinaAPI.",
    "deployment_inference": "ChinaAPI fit assessment inferred from modality, model size, context and documented serving path; not a quality benchmark."
  },
  "categories": {
    "general_agent": "General and agentic foundation models",
    "world_model": "Digital-world and environment models",
    "robotics_vla": "Physical-world vision-language-action models",
    "audio_video_generation": "Joint audio-video generation",
    "speech_recognition": "Automatic speech recognition"
  },
  "workload_profiles": {
    "minimum": "One request, short context, no SLA; proves runnable, not production-ready.",
    "team_200": "200 seats, 80-120 DAU, 20-30 peak online, 8-16 concurrent generations, 8K-32K ordinary prompts and occasional 128K agent work.",
    "commercial_api": "Public multi-tenant API, redundancy, rolling updates, rate limits and at least 99.9% availability target. Final GPU count requires measured traffic and latency targets."
  },
  "models": [
    {
      "id": "moonshotai/kimi-k3",
      "name": "Kimi K3",
      "vendor": "Moonshot AI",
      "release_date": "2026-07-27",
      "license": "Kimi K3 License",
      "commercial_use": {
        "status": "conditional",
        "maas": "Separate agreement required when a MaaS operator and affiliates exceed USD 20M aggregate revenue over any consecutive 12 months.",
        "attribution": "Display Kimi K3 prominently above 100M MAU or USD 20M monthly revenue.",
        "exceptions": "The cited requirements do not apply to internal use or use through Moonshot official products or certified inference partners.",
        "license_url": "https://github.com/MoonshotAI/Kimi-K3/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "2.8T",
      "parameters_active": "104B",
      "native_precision": "MXFP4 routed experts; BF16 dense components; MXFP8 activations",
      "context_window": 1048576,
      "frameworks": [
        "vLLM",
        "SGLang",
        "TokenSpeed"
      ],
      "weight_url": "https://huggingface.co/moonshotai/Kimi-K3",
      "official_url": "https://github.com/MoonshotAI/Kimi-K3",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image",
          "video"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Frontier-scale long-context multimodal and coding agent workloads where a cluster deployment is acceptable.",
      "scenarios": [
        {
          "name": "Long-context coding agents",
          "evidence": "vendor_positioning",
          "finding": "Positioned for repository-scale coding and long-horizon agent work.",
          "source_url": "https://github.com/MoonshotAI/Kimi-K3"
        },
        {
          "name": "Multimodal document and video analysis",
          "evidence": "vendor_positioning",
          "finding": "Official materials expose native multimodal inputs and a 1M-token context window.",
          "source_url": "https://github.com/MoonshotAI/Kimi-K3"
        },
        {
          "name": "High-value enterprise agent workflows",
          "evidence": "deployment_inference",
          "finding": "Best considered when model capability justifies data-center-class full-context workers."
        }
      ],
      "not_recommended": [
        "Single-workstation deployments",
        "Low-cost high-QPS chat without aggressive batching or distillation"
      ],
      "tiers": {
        "minimum": {
          "status": "not_published",
          "evidence": "C",
          "finding": "Vendor documents supported engines but does not publish a smallest runnable GPU configuration."
        },
        "team_200": {
          "status": "requires_cluster_sizing",
          "evidence": "E",
          "finding": "Not responsibly specifiable before a concurrency and context benchmark; workstation deployment is not credible."
        },
        "commercial_api": {
          "status": "validated_reference",
          "evidence": "B",
          "finding": "NVIDIA Dynamo publishes full-1M profiles using 8x GB300 or 16x GB200 per aggregated worker; disaggregated profiles use more GPUs.",
          "source_url": "https://github.com/ai-dynamo/dynamo/blob/main/recipes/kimi-k3/README.md"
        }
      },
      "limitations": [
        "No ChinaAPI hardware reproduction",
        "Full-context reference configurations are data-center cluster class"
      ]
    },
    {
      "id": "zai-org/glm-5.2",
      "name": "GLM-5.2",
      "vendor": "Z.ai",
      "release_date": "2026-07-12",
      "license": "MIT",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in the MIT license.",
        "attribution": "Retain the copyright and permission notice.",
        "license_url": "https://huggingface.co/zai-org/GLM-5.2/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "753B",
      "parameters_active": "not independently verified in V1",
      "native_precision": "BF16 and official FP8 checkpoint",
      "context_window": 1048576,
      "frameworks": [
        "vLLM",
        "SGLang",
        "xLLM",
        "KTransformers"
      ],
      "weight_url": "https://huggingface.co/zai-org/GLM-5.2",
      "official_url": "https://github.com/zai-org/GLM-5",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Long-horizon software engineering, tool use and million-token analysis on enterprise infrastructure.",
      "scenarios": [
        {
          "name": "Long-horizon software engineering",
          "evidence": "vendor_positioning",
          "finding": "Official release materials emphasize coding-agent and tool-use workloads.",
          "source_url": "https://github.com/zai-org/GLM-5"
        },
        {
          "name": "Million-token repository and document analysis",
          "evidence": "vendor_positioning",
          "finding": "The official checkpoint supports a 1M-token context window.",
          "source_url": "https://huggingface.co/zai-org/GLM-5.2"
        },
        {
          "name": "Private enterprise agent service",
          "evidence": "deployment_inference",
          "finding": "A credible fit when CPU/GPU offload or TP8 infrastructure is already available."
        }
      ],
      "not_recommended": [
        "Latency-sensitive single-GPU serving",
        "Capacity planning based only on employee count"
      ],
      "tiers": {
        "minimum": {
          "status": "heterogeneous_reference",
          "evidence": "B",
          "finding": "KTransformers documents an 8-GPU CPU/GPU-offload launch shape with 96 CPU inference threads; this is a tutorial target, not a claimed absolute minimum.",
          "source_url": "https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/kt-kernel/GLM-5.2-Tutorial.md"
        },
        "team_200": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "Start from a replicated TP8 service only after measuring the target context mix; no official 200-seat capacity claim exists."
        },
        "commercial_api": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "Requires redundant replicas or disaggregated prefill/decode; GPU count cannot be inferred from seats alone."
        }
      },
      "limitations": [
        "The 8-GPU tutorial is not an official minimum claim",
        "No public V1 concurrency result"
      ]
    },
    {
      "id": "deepseek-ai/deepseek-v4-pro",
      "name": "DeepSeek V4 Pro",
      "vendor": "DeepSeek",
      "release_date": "2026-06-22",
      "license": "MIT",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in the MIT license.",
        "attribution": "Retain the copyright and permission notice.",
        "license_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "862B reported by model repository",
      "parameters_active": "not independently verified in V1",
      "native_precision": "mixed-precision checkpoint, approximately 960 GB in vLLM recipe",
      "context_window": 1048576,
      "frameworks": [
        "vLLM",
        "SGLang"
      ],
      "weight_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
      "official_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Maximum-quality reasoning, coding and long-context agents when an eight-accelerator worker is viable.",
      "scenarios": [
        {
          "name": "Complex reasoning and coding",
          "evidence": "public_benchmark",
          "finding": "The official model card publishes reasoning, coding and agent benchmark results.",
          "source_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
        },
        {
          "name": "Million-token retrieval and analysis",
          "evidence": "vendor_positioning",
          "finding": "The release is designed around efficient million-token context intelligence.",
          "source_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro"
        },
        {
          "name": "Premium private agent endpoint",
          "evidence": "deployment_inference",
          "finding": "Use when quality has more value than minimum infrastructure cost."
        }
      ],
      "not_recommended": [
        "Budget workstation inference",
        "Public API launch without at least one redundant worker"
      ],
      "tiers": {
        "minimum": {
          "status": "validated_reference_not_minimum",
          "evidence": "B",
          "finding": "vLLM publishes 8x B300, 8x H200, 8x MI355X and GB200 profiles; none is labelled the absolute minimum.",
          "source_url": "https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4-Pro.yaml"
        },
        "team_200": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "An 8-GPU worker may be a capacity building block, but replicas depend on concurrency and output-token demand."
        },
        "commercial_api": {
          "status": "validated_reference",
          "evidence": "B",
          "finding": "8x H200 is documented with context capped at 800K to preserve KV headroom; production redundancy requires additional workers.",
          "source_url": "https://github.com/vllm-project/recipes/blob/main/models/deepseek-ai/DeepSeek-V4-Pro.yaml"
        }
      },
      "limitations": [
        "Some SM120 workstation kernels have unresolved reports",
        "Loading weights is not proof that the first forward pass succeeds"
      ]
    },
    {
      "id": "minimaxai/minimax-m3",
      "name": "MiniMax M3",
      "vendor": "MiniMax",
      "release_date": "2026-06-01",
      "license": "MiniMax Community License",
      "commercial_use": {
        "status": "conditional",
        "maas": "Commercial API and hosted use are Commercial Use. Above USD 20M yearly revenue obtain prior written authorization; otherwise send the required one-time notice.",
        "attribution": "Prominently display Built with MiniMax M3 for commercial use.",
        "prohibited_uses": "The license includes specified unlawful, military and harmful-use restrictions.",
        "license_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "approximately 428B",
      "parameters_active": "approximately 23B",
      "native_precision": "BF16 plus official MXFP8; NVIDIA NVFP4 conversion available",
      "context_window": 1048576,
      "frameworks": [
        "Transformers",
        "vLLM nightly",
        "SGLang"
      ],
      "weight_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3",
      "official_url": "https://github.com/MiniMax-AI/MiniMax-M3",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image",
          "video"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Multimodal agent, coding and long-context workloads with explicit commercial-license review.",
      "scenarios": [
        {
          "name": "Multimodal agent workflows",
          "evidence": "vendor_positioning",
          "finding": "Official materials position M3 for multimodal perception and agent tasks.",
          "source_url": "https://github.com/MiniMax-AI/MiniMax-M3"
        },
        {
          "name": "Coding and tool orchestration",
          "evidence": "public_benchmark",
          "finding": "The official release reports coding and agent evaluations.",
          "source_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3"
        },
        {
          "name": "Commercial embedded assistant",
          "evidence": "deployment_inference",
          "finding": "Potentially suitable after the revenue, notice and attribution clauses are cleared."
        }
      ],
      "not_recommended": [
        "Commercial launch before license notice and attribution review",
        "Stable-production serving without pinned framework versions"
      ],
      "tiers": {
        "minimum": {
          "status": "candidate_recipe_under_review",
          "evidence": "D",
          "finding": "A 4x RTX PRO 6000 NVFP4 recipe is under review upstream; V1 does not promote an unmerged recipe to validated minimum."
        },
        "team_200": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "A quantized multi-GPU worker is plausible, but the 200-seat recommendation needs measured TTFT, throughput and context mix."
        },
        "commercial_api": {
          "status": "framework_support_maturing",
          "evidence": "B",
          "finding": "Aggregated and disaggregated vLLM recipes are still landing; pinning a nightly build may be required."
        }
      },
      "limitations": [
        "Stable vLLM release support was not complete at the V1 cutoff",
        "Do not equate a pending recipe with successful ChinaAPI reproduction"
      ]
    },
    {
      "id": "qwen/qwen3.6-35b-a3b",
      "name": "Qwen3.6-35B-A3B",
      "vendor": "Alibaba Qwen",
      "release_date": "2026-04-16",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and required notices, preserve attribution notices, and mark modified files; trademark rights are not granted.",
        "license_url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "35B",
      "parameters_active": "3B",
      "native_precision": "BF16; official and community quantized formats available",
      "context_window": 262144,
      "frameworks": [
        "Transformers",
        "vLLM",
        "SGLang",
        "llama.cpp",
        "MLX"
      ],
      "weight_url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B",
      "official_url": "https://github.com/QwenLM/Qwen3.6",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image",
          "video"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "A comparatively deployable private multimodal agent baseline with strong ecosystem coverage.",
      "scenarios": [
        {
          "name": "Private coding and office assistant",
          "evidence": "deployment_inference",
          "finding": "The 35B/3B-active shape and broad serving support make it a practical baseline for controlled workloads."
        },
        {
          "name": "Visual document and UI understanding",
          "evidence": "vendor_positioning",
          "finding": "Qwen3.6 is released as a native multimodal model family.",
          "source_url": "https://github.com/QwenLM/Qwen3.6"
        },
        {
          "name": "Tool-calling agent service",
          "evidence": "vendor_positioning",
          "finding": "Official serving examples include tool parsing and long-context operation.",
          "source_url": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B"
        }
      ],
      "not_recommended": [
        "Assuming TP4 equals four production replicas",
        "Unvalidated parser upgrades in a critical tool loop"
      ],
      "tiers": {
        "minimum": {
          "status": "official_launch_shape",
          "evidence": "C",
          "finding": "Official serving examples use TP4 and 262K context. Smaller quantized short-context shapes exist, but V1 does not call them the official minimum."
        },
        "team_200": {
          "status": "best_v1_benchmark_candidate",
          "evidence": "E",
          "finding": "Use two measured serving replicas as the initial HA design candidate; exact GPUs remain pending load tests."
        },
        "commercial_api": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "Scale through replicated workers after measuring prefill-heavy and decode-heavy traffic separately."
        }
      },
      "limitations": [
        "Official TP4 example is a launch shape, not a capacity guarantee",
        "Tool-call parser issues have been reported for related Qwen3.5 configurations"
      ]
    },
    {
      "id": "deepseek-ai/deepseek-v4-flash",
      "name": "DeepSeek V4 Flash",
      "vendor": "DeepSeek",
      "release_date": "2026-06-22",
      "license": "MIT",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in the MIT license.",
        "attribution": "Retain the copyright and permission notice.",
        "license_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "284B",
      "parameters_active": "13B",
      "native_precision": "FP4 experts with FP8 dense components",
      "context_window": 1048576,
      "frameworks": [
        "Transformers",
        "vLLM",
        "SGLang"
      ],
      "weight_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
      "official_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "A smaller DeepSeek V4 worker for high-frequency coding, reasoning and agent traffic.",
      "scenarios": [
        {
          "name": "High-frequency coding agents",
          "evidence": "public_benchmark",
          "finding": "The official card reports coding and agent results and compares Flash reasoning modes.",
          "source_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
        },
        {
          "name": "Million-token analysis",
          "evidence": "vendor_positioning",
          "finding": "The model supports a 1M-token context with a 284B/13B-active MoE shape.",
          "source_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
        },
        {
          "name": "Throughput-sensitive private endpoint",
          "evidence": "deployment_inference",
          "finding": "Prefer over V4 Pro when infrastructure cost and concurrency matter more than maximum quality."
        }
      ],
      "not_recommended": [
        "Treating Flash as quality-equivalent to Pro on every task",
        "Launching a public API without measured parser and long-context behavior"
      ],
      "tiers": {
        "minimum": {
          "status": "vendor_run_path_not_minimum",
          "evidence": "C",
          "finding": "Official local instructions and engine integrations are published, but no absolute minimum GPU count is claimed.",
          "source_url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash"
        },
        "team_200": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "Use one measured tensor-parallel worker as the capacity baseline; add replicas only after concurrency tests."
        },
        "commercial_api": {
          "status": "framework_supported",
          "evidence": "B",
          "finding": "vLLM and SGLang support are available; production still requires redundant workers and pinned encoding logic.",
          "source_url": "https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-04-24-deepseek-v4.md"
        }
      },
      "limitations": [
        "No ChinaAPI hardware reproduction",
        "The release uses a dedicated encoding implementation rather than a Jinja chat template"
      ]
    },
    {
      "id": "stepfun-ai/step-3.7-flash",
      "name": "Step 3.7 Flash",
      "vendor": "StepFun",
      "release_date": "2026-05-28",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and notices, preserve attribution notices, and mark modified files.",
        "license_url": "https://github.com/stepfun-ai/Step-3.7-Flash/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "198B",
      "parameters_active": "approximately 11B",
      "native_precision": "BF16, FP8, NVFP4 and GGUF releases",
      "context_window": 262144,
      "frameworks": [
        "vLLM",
        "SGLang",
        "Transformers",
        "llama.cpp"
      ],
      "weight_url": "https://huggingface.co/stepfun-ai/Step-3.7-Flash",
      "official_url": "https://github.com/stepfun-ai/Step-3.7-Flash",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "High-frequency multimodal agents, financial-document parsing, verified search loops and concurrent coding.",
      "scenarios": [
        {
          "name": "Visual documents and UI-to-code",
          "evidence": "vendor_positioning",
          "finding": "Official materials highlight charts, GUIs, wireframes and structured-code extraction.",
          "source_url": "https://github.com/stepfun-ai/Step-3.7-Flash"
        },
        {
          "name": "Search and tool orchestration",
          "evidence": "public_benchmark",
          "finding": "The vendor publishes ClawEval, Toolathlon and tool-use evaluations.",
          "source_url": "https://github.com/stepfun-ai/Step-3.7-Flash"
        },
        {
          "name": "Concurrent coding agents",
          "evidence": "vendor_positioning",
          "finding": "The model is explicitly engineered for high-frequency production agent workloads.",
          "source_url": "https://github.com/stepfun-ai/Step-3.7-Flash"
        }
      ],
      "not_recommended": [
        "Using the 128GB local path as a production throughput claim",
        "Assuming vendor benchmark throughput transfers to long-context prefill"
      ],
      "tiers": {
        "minimum": {
          "status": "official_local_minimum",
          "evidence": "C",
          "finding": "The official GGUF path specifies about 120GB minimum unified memory/VRAM and recommends 128GB.",
          "source_url": "https://github.com/stepfun-ai/Step-3.7-Flash"
        },
        "team_200": {
          "status": "production_candidate",
          "evidence": "E",
          "finding": "Start with one TP4 NVFP4 or TP8 FP8 worker and benchmark the actual image and context mix."
        },
        "commercial_api": {
          "status": "official_launch_shapes",
          "evidence": "C",
          "finding": "Official examples publish TP4 NVFP4 and TP8 FP8/BF16 serving shapes; replicas are still required for HA.",
          "source_url": "https://github.com/stepfun-ai/Step-3.7-Flash"
        }
      },
      "limitations": [
        "Published 400 tok/s is a vendor benchmark, not a universal SLA",
        "NVFP4 support depends on recent engine and GPU paths"
      ]
    },
    {
      "id": "meituan-longcat/longcat-2.0",
      "name": "LongCat 2.0",
      "vendor": "Meituan",
      "release_date": "2026-07-02",
      "license": "MIT",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in the MIT license.",
        "attribution": "Retain the copyright and permission notice.",
        "license_url": "https://github.com/meituan-longcat/LongCat-2.0/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "1.6T",
      "parameters_active": "approximately 48B",
      "native_precision": "Vendor checkpoint; exact serving precision varies by recipe",
      "context_window": 1048576,
      "frameworks": [
        "SGLang",
        "SGLang-FluentLLM"
      ],
      "weight_url": "https://huggingface.co/meituan-longcat/LongCat-2.0",
      "official_url": "https://github.com/meituan-longcat/LongCat-2.0",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Long-horizon coding, search, repository edits and tool-driven agents on GPU or NPU clusters.",
      "scenarios": [
        {
          "name": "Repository-scale coding",
          "evidence": "vendor_positioning",
          "finding": "Official materials emphasize repository edits and integrations with coding-agent harnesses.",
          "source_url": "https://github.com/meituan-longcat/LongCat-2.0"
        },
        {
          "name": "Search and general agents",
          "evidence": "public_benchmark",
          "finding": "The vendor publishes coding, BrowseComp, RWSearch and agent evaluations.",
          "source_url": "https://github.com/meituan-longcat/LongCat-2.0"
        },
        {
          "name": "NPU-based sovereign deployment",
          "evidence": "vendor_positioning",
          "finding": "An official SGLang-FluentLLM NPU serving path is linked.",
          "source_url": "https://github.com/meituan-longcat/LongCat-2.0"
        }
      ],
      "not_recommended": [
        "Workstation deployment",
        "Quoting in-house benchmark results as ChinaAPI reproduction"
      ],
      "tiers": {
        "minimum": {
          "status": "not_published",
          "evidence": "C",
          "finding": "GPU and NPU paths are documented, but the vendor does not publish an absolute minimum hardware shape."
        },
        "team_200": {
          "status": "cluster_sizing_required",
          "evidence": "E",
          "finding": "A 1.6T model needs a measured cluster design; employee count alone is not a capacity input."
        },
        "commercial_api": {
          "status": "official_serving_paths",
          "evidence": "C",
          "finding": "Official GPU and NPU serving paths exist; topology, redundancy and throughput remain operator-specific.",
          "source_url": "https://github.com/meituan-longcat/LongCat-2.0"
        }
      },
      "limitations": [
        "No official minimum GPU count",
        "Most published evaluation values are vendor-measured"
      ]
    },
    {
      "id": "xiaomimimo/mimo-v2.5",
      "name": "MiMo V2.5",
      "vendor": "Xiaomi MiMo",
      "release_date": "2026-04-27",
      "license": "MIT",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in the MIT license.",
        "attribution": "Retain the copyright and permission notice.",
        "license_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "310B",
      "parameters_active": "15B",
      "native_precision": "BF16 and FP8 components",
      "context_window": 1048576,
      "frameworks": [
        "Transformers",
        "SGLang",
        "vLLM"
      ],
      "weight_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5",
      "official_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image",
          "video",
          "audio"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "Native omnimodal understanding, long-context reasoning and agentic workflows across text, image, video and audio.",
      "scenarios": [
        {
          "name": "Omnimodal research and support agents",
          "evidence": "vendor_positioning",
          "finding": "The checkpoint natively accepts text, image, video and audio.",
          "source_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5"
        },
        {
          "name": "Long video, audio and document analysis",
          "evidence": "vendor_positioning",
          "finding": "The model supports up to 1M context and dedicated visual and audio encoders.",
          "source_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5"
        },
        {
          "name": "Multimodal tool-using agents",
          "evidence": "public_benchmark",
          "finding": "The official card publishes multimodal, coding, agent and long-context evaluations.",
          "source_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5"
        }
      ],
      "not_recommended": [
        "Small single-GPU deployment",
        "Using stale config or tokenizer files from the initial release"
      ],
      "tiers": {
        "minimum": {
          "status": "not_published",
          "evidence": "C",
          "finding": "Transformers can load the checkpoint, but the vendor does not state an absolute minimum GPU configuration."
        },
        "team_200": {
          "status": "benchmark_required",
          "evidence": "E",
          "finding": "Use the official distributed recipe as a starting worker and size replicas from the real modality mix."
        },
        "commercial_api": {
          "status": "official_distributed_recipe",
          "evidence": "C",
          "finding": "The official card shows an FP8 SGLang DP2×TP8 configuration at 262K context.",
          "source_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5"
        }
      },
      "limitations": [
        "Official deployment example is not a minimum",
        "Audio and video traffic need separate encoder-capacity measurements"
      ]
    },
    {
      "id": "qwen/qwen3.5-9b",
      "name": "Qwen3.5-9B",
      "vendor": "Alibaba Qwen",
      "release_date": "2026-02-16",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and notices, preserve attribution notices, and mark modified files.",
        "license_url": "https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "9B",
      "parameters_active": "9B",
      "native_precision": "BF16; multiple quantized formats available",
      "context_window": 262144,
      "frameworks": [
        "Transformers",
        "vLLM",
        "SGLang",
        "KTransformers",
        "llama.cpp"
      ],
      "weight_url": "https://huggingface.co/Qwen/Qwen3.5-9B",
      "official_url": "https://huggingface.co/Qwen/Qwen3.5-9B",
      "category": "general_agent",
      "modalities": {
        "input": [
          "text",
          "image",
          "video"
        ],
        "output": [
          "text"
        ]
      },
      "scenario_summary": "The local and edge-friendly baseline for private multimodal assistants, extraction and moderate-volume APIs.",
      "scenarios": [
        {
          "name": "Local multimodal assistant",
          "evidence": "deployment_inference",
          "finding": "The 9B dense shape is the most accessible formal model in this ledger."
        },
        {
          "name": "Visual extraction and classification",
          "evidence": "vendor_positioning",
          "finding": "The official model is a unified vision-language foundation model.",
          "source_url": "https://huggingface.co/Qwen/Qwen3.5-9B"
        },
        {
          "name": "Cost-sensitive private API",
          "evidence": "deployment_inference",
          "finding": "A better first benchmark target than frontier-scale MoE models when concurrency and budget dominate."
        }
      ],
      "not_recommended": [
        "Assuming 1M extended context fits a 24GB card",
        "Highest-complexity long-horizon agents without task-specific evaluation"
      ],
      "tiers": {
        "minimum": {
          "status": "memory_estimate_not_official",
          "evidence": "E",
          "finding": "A 24GB-class GPU is a reasonable BF16 short-context planning target, but the vendor does not label this an official minimum."
        },
        "team_200": {
          "status": "best_low_cost_candidate",
          "evidence": "E",
          "finding": "Benchmark one or two 24–48GB workers before considering larger models; exact replicas depend on output length."
        },
        "commercial_api": {
          "status": "mainstream_framework_support",
          "evidence": "C",
          "finding": "Official Transformers, vLLM and SGLang examples are published; add redundant replicas for availability.",
          "source_url": "https://huggingface.co/Qwen/Qwen3.5-9B"
        }
      },
      "limitations": [
        "24GB guidance is an estimate, not ChinaAPI reproduction",
        "Native 262K and extended 1M contexts materially increase KV-cache demand"
      ]
    },
    {
      "id": "qwen/qwen-agentworld-35b-a3b",
      "name": "Qwen-AgentWorld-35B-A3B",
      "vendor": "Alibaba Qwen",
      "release_date": "2026-06-24",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and notices, preserve attribution notices, and mark modified files.",
        "license_url": "https://github.com/QwenLM/Qwen-AgentWorld/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "35B",
      "parameters_active": "3B",
      "native_precision": "BF16",
      "context_window": 262144,
      "frameworks": [
        "SGLang",
        "vLLM"
      ],
      "weight_url": "https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B",
      "official_url": "https://github.com/QwenLM/Qwen-AgentWorld",
      "category": "world_model",
      "modalities": {
        "input": [
          "text",
          "environment state",
          "action history"
        ],
        "output": [
          "predicted environment state"
        ]
      },
      "scenario_summary": "A language world model for simulating MCP, Search, Terminal, SWE, Android, Web and OS agent environments.",
      "scenarios": [
        {
          "name": "Agent environment simulation",
          "evidence": "vendor_positioning",
          "finding": "The model predicts environment transitions across seven unified domains.",
          "source_url": "https://github.com/QwenLM/Qwen-AgentWorld"
        },
        {
          "name": "Synthetic trajectories and perturbation tests",
          "evidence": "vendor_positioning",
          "finding": "Official materials highlight controllable simulation and fictional-world construction.",
          "source_url": "https://github.com/QwenLM/Qwen-AgentWorld"
        },
        {
          "name": "Agent regression evaluation",
          "evidence": "deployment_inference",
          "finding": "Useful as a simulator component, not as a drop-in customer chatbot."
        }
      ],
      "not_recommended": [
        "General-purpose chat replacement",
        "Treating simulated success as proof of real-environment reliability"
      ],
      "tiers": {
        "minimum": {
          "status": "official_tp4_launch",
          "evidence": "C",
          "finding": "Official SGLang and vLLM examples use tensor parallel size 4.",
          "source_url": "https://github.com/QwenLM/Qwen-AgentWorld"
        },
        "team_200": {
          "status": "not_a_seat_based_service",
          "evidence": "E",
          "finding": "Size by simulation jobs and trajectory length, not employee seats."
        },
        "commercial_api": {
          "status": "specialized_service",
          "evidence": "E",
          "finding": "Expose behind a task-specific simulator contract; do not market it as a normal chat-completions quality substitute."
        }
      },
      "limitations": [
        "World-model outputs are simulations",
        "The 35B release covers seven named domains, not arbitrary physical environments"
      ]
    },
    {
      "id": "xiaomirobotics/xiaomi-robotics-0",
      "name": "Xiaomi-Robotics-0",
      "vendor": "Xiaomi Robotics",
      "release_date": "2026-02-11",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and notices, preserve attribution notices, and mark modified files.",
        "license_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "4.7B",
      "parameters_active": "4.7B",
      "native_precision": "BF16 with Flash Attention 2",
      "context_window": 0,
      "frameworks": [
        "Transformers",
        "PyTorch"
      ],
      "weight_url": "https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-0",
      "official_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0",
      "category": "robotics_vla",
      "modalities": {
        "input": [
          "robot camera",
          "language instruction",
          "proprioception"
        ],
        "output": [
          "robot action"
        ]
      },
      "scenario_summary": "Real-time robotic manipulation research and post-training across supported embodiments and simulation suites.",
      "scenarios": [
        {
          "name": "Robot manipulation",
          "evidence": "vendor_positioning",
          "finding": "The official deployment guide targets robotic manipulation with asynchronous real-time execution.",
          "source_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0"
        },
        {
          "name": "LIBERO, CALVIN and SimplerEnv evaluation",
          "evidence": "public_benchmark",
          "finding": "Official fine-tuned checkpoints and evaluation guides are released for all three suites.",
          "source_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0"
        },
        {
          "name": "Embodiment-specific post-training",
          "evidence": "vendor_positioning",
          "finding": "Post-training code is available for adapting the model to new data.",
          "source_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0"
        }
      ],
      "not_recommended": [
        "Direct deployment on an unvalidated physical robot",
        "Safety-critical control without independent interlocks and task-specific validation"
      ],
      "tiers": {
        "minimum": {
          "status": "consumer_gpu_claim_no_vram",
          "evidence": "C",
          "finding": "The vendor says BF16 inference is optimized for consumer GPUs but publishes no exact minimum VRAM.",
          "source_url": "https://github.com/XiaomiRobotics/Xiaomi-Robotics-0"
        },
        "team_200": {
          "status": "robot_fleet_profile_required",
          "evidence": "E",
          "finding": "Size by robots, camera rate and control latency; a 200-seat office profile is not applicable."
        },
        "commercial_api": {
          "status": "edge_control_not_public_api",
          "evidence": "E",
          "finding": "Prefer an on-robot or near-edge safety architecture; a remote multi-tenant API is not the default production shape."
        }
      },
      "limitations": [
        "Simulation benchmarks do not prove physical-world safety",
        "Exact VRAM and end-to-end control latency are not published"
      ]
    },
    {
      "id": "baidu/nava",
      "name": "NAVA",
      "vendor": "Baidu ERNIE Team",
      "release_date": "2026-05-28",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "conditional",
        "maas": "The model card states Apache-2.0, but bundled LTX audio-VAE artifacts carry an additional community license that must be reviewed for the shipped stack.",
        "attribution": "Preserve Apache notices and the notices/licenses for bundled upstream components.",
        "prohibited_uses": "The model card prohibits depicting real persons without consent, including face or voice likeness reproduction.",
        "license_url": "https://huggingface.co/baidu/NAVA/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "6.3B backbone",
      "parameters_active": "6.3B",
      "native_precision": "BF16",
      "context_window": 0,
      "frameworks": [
        "PyTorch",
        "Ulysses sequence parallel"
      ],
      "weight_url": "https://huggingface.co/baidu/NAVA",
      "official_url": "https://github.com/ernie-research/NAVA",
      "category": "audio_video_generation",
      "modalities": {
        "input": [
          "text",
          "image",
          "reference voice"
        ],
        "output": [
          "video",
          "stereo audio",
          "speech"
        ]
      },
      "scenario_summary": "Synchronized audio-video generation with multi-speaker timbre control, camera prompting and image continuation.",
      "scenarios": [
        {
          "name": "Synchronized short-form audio-video",
          "evidence": "vendor_positioning",
          "finding": "NAVA jointly generates video, scene audio and speech rather than aligning separate outputs after generation.",
          "source_url": "https://huggingface.co/baidu/NAVA"
        },
        {
          "name": "Multi-speaker and reference-timbre scenes",
          "evidence": "vendor_positioning",
          "finding": "The official checkpoint supports up to two reference voices bound to speech spans.",
          "source_url": "https://huggingface.co/baidu/NAVA"
        },
        {
          "name": "720p creative generation",
          "evidence": "public_benchmark",
          "finding": "The vendor reports VerseBench synchronization and quality results plus an 8-GPU fast path.",
          "source_url": "https://huggingface.co/baidu/NAVA"
        }
      ],
      "not_recommended": [
        "Unconsented face or voice cloning",
        "Low-latency interactive video generation"
      ],
      "tiers": {
        "minimum": {
          "status": "single_gpu_supported_no_vram",
          "evidence": "C",
          "finding": "The model card supports single-GPU inference but does not publish exact minimum VRAM.",
          "source_url": "https://huggingface.co/baidu/NAVA"
        },
        "team_200": {
          "status": "media_queue_required",
          "evidence": "E",
          "finding": "Size a queued render farm by jobs per hour, resolution and duration; office-seat assumptions do not apply."
        },
        "commercial_api": {
          "status": "official_8gpu_reference",
          "evidence": "C",
          "finding": "The official Ulysses SP8 path reports roughly one minute for a 720p synchronized clip; HA needs additional workers.",
          "source_url": "https://huggingface.co/baidu/NAVA"
        }
      },
      "limitations": [
        "Default clips are about 6–10 seconds",
        "The full dependency stack includes component-specific notices beyond the headline Apache license"
      ]
    },
    {
      "id": "xiaomimimo/mimo-v2.5-asr",
      "name": "MiMo-V2.5-ASR",
      "vendor": "Xiaomi MiMo",
      "release_date": "2026-04-23",
      "license": "Apache-2.0",
      "commercial_use": {
        "status": "permitted",
        "maas": "No model-specific MaaS restriction identified in Apache-2.0.",
        "attribution": "Provide the license and notices, preserve attribution notices, and mark modified files.",
        "license_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR/blob/main/LICENSE",
        "review_status": "primary_source_reviewed_not_legal_advice"
      },
      "parameters_total": "not published",
      "parameters_active": "not published",
      "native_precision": "CUDA 12 path with Flash Attention",
      "context_window": 0,
      "frameworks": [
        "PyTorch",
        "Gradio"
      ],
      "weight_url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR",
      "official_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR",
      "category": "speech_recognition",
      "modalities": {
        "input": [
          "audio"
        ],
        "output": [
          "punctuated text"
        ]
      },
      "scenario_summary": "Chinese and English transcription across dialects, code-switching, meetings, songs, noise and knowledge-heavy speech.",
      "scenarios": [
        {
          "name": "Chinese dialect and code-switch transcription",
          "evidence": "vendor_positioning",
          "finding": "Official support includes Wu, Cantonese, Hokkien, Sichuanese and Chinese-English code switching.",
          "source_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR"
        },
        {
          "name": "Meetings and noisy far-field audio",
          "evidence": "vendor_positioning",
          "finding": "The release targets overlapping speakers, heavy noise and far-field capture.",
          "source_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR"
        },
        {
          "name": "Knowledge-heavy and lyric transcription",
          "evidence": "public_benchmark",
          "finding": "The vendor reports evaluations across dialects, lyrics and complex English scenarios.",
          "source_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR"
        }
      ],
      "not_recommended": [
        "Capacity promises before real-time-factor and batch testing",
        "Assuming speaker overlap performance replaces diarization requirements"
      ],
      "tiers": {
        "minimum": {
          "status": "not_published",
          "evidence": "C",
          "finding": "The official local Gradio and Python path requires CUDA 12+, but exact minimum VRAM is not published.",
          "source_url": "https://github.com/XiaomiMiMo/MiMo-V2.5-ASR"
        },
        "team_200": {
          "status": "audio_workload_profile_required",
          "evidence": "E",
          "finding": "Size by audio hours, peak simultaneous streams and latency rather than office seats."
        },
        "commercial_api": {
          "status": "custom_service_required",
          "evidence": "E",
          "finding": "The release provides local inference code, not a validated multi-tenant serving recipe; build queueing, batching and observability."
        }
      },
      "limitations": [
        "Model parameter count and minimum VRAM are not published",
        "The public release does not provide an official high-throughput serving benchmark"
      ]
    }
  ],
  "watchlist": [
    {
      "name": "ByteDance Lance",
      "status": "license_unclear",
      "reason": "Official weights and inference code exist, but the repository does not clearly name a standard model license; commercial rights need clarification.",
      "source_url": "https://github.com/bytedance/Lance"
    },
    {
      "name": "GigaWorld-1",
      "status": "partial_release",
      "reason": "Official code and selected weights are available as a research preview, but the complete deployment surface is not yet equivalent to the formal ledger.",
      "source_url": "https://github.com/open-gigaai/giga-world-1"
    },
    {
      "name": "Qwen-VLA",
      "status": "weights_not_confirmed",
      "reason": "The official repository is public, but a complete, directly deployable released checkpoint was not sufficiently confirmed at this cutoff.",
      "source_url": "https://github.com/QwenLM/Qwen-VLA"
    }
  ],
  "canonical": "https://chinaapi.ai/open-model-deployment/",
  "methodology_url": "https://chinaapi.ai/open-model-deployment/#methodology",
  "license_note": "Model-license summaries are product research, not legal advice. Review the linked license text and obtain counsel for a commercial launch."
}